A Collaborative Filtering Recommendation Method and System Based on State-Space Model
By using a collaborative filtering method based on a state-space model and leveraging the Mamba state-space model to process user interaction data, this approach addresses the problem of insufficient understanding of user preferences in traditional collaborative filtering methods, achieving more accurate recommendation results.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- JIMEI UNIV
- Filing Date
- 2026-01-29
- Publication Date
- 2026-05-26
AI Technical Summary
Traditional collaborative filtering methods rely on only a single user-item pair, which leads to an inability to fully understand user preferences, fragmented interaction context, loss of sequence preference information, and an inability to accurately predict user preferences.
A collaborative filtering method based on a state-space model is adopted. By obtaining the item and user embedding representations of the user interaction item set, the input sequence is constructed, and the Mamba state-space model is used for encoding to generate an encoded vector to predict the interaction score between the user and the candidate items.
By effectively capturing the global interaction history between users and items, avoiding fragmented interaction contexts, the accuracy and precision of recommendations are improved, and the overall performance of the recommendation system is enhanced.
Smart Images
Figure CN121597919B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of recommendation system technology, and in particular relates to a collaborative filtering recommendation method and system based on a state-space model. Background Technology
[0002] In today's digital age, recommendation systems have become a core component for enhancing user experience across various platforms, with applications spanning e-commerce websites, streaming services, and social media platforms. These systems aim to provide users with personalized suggestions, thereby increasing user engagement and satisfaction, and ultimately enhancing the overall performance and revenue of the platform.
[0003] The core challenge of such systems lies in predicting user preferences based on their historical behavior. The mainstream solution is collaborative filtering (CF). CF has become one of the most influential technologies, generating recommendations solely based on the user-item interaction matrix, without requiring explicit user profiles or content-side information. The prerequisite for applying CF is the assumption that "users with similar past behaviors will have similar future preferences."
[0004] Traditional CF (Conceptual Filtering) methods are widely used, but they face a major challenge: using only (user, item) pairs as training samples leads to an inability to fully understand user preferences, fragmented interaction context, and loss of sequence preference information. Therefore, traditional CF methods struggle to achieve accurate predictions. Summary of the Invention
[0005] To address the aforementioned technical problems, this invention proposes a collaborative filtering recommendation method and system based on a state-space model. This invention aims to solve the problem that traditional CF methods rely only on a single user-item pair, which leads to the inability of CF methods to accurately predict user preferences.
[0006] To achieve the above objectives, this invention provides a collaborative filtering recommendation method based on a state-space model, comprising:
[0007] Based on the user's item interaction data, obtain the set of items the user interacted with;
[0008] Based on the identification information of each item in the user interaction item set and the user's identification information, obtain the item embedding representation and user embedding representation of each item respectively;
[0009] Construct the input sequence based on the item embedding representation and the user embedding representation;
[0010] Based on the input sequence, an encoding vector is obtained by encoding through a state-space model.
[0011] Based on the encoded vector, the user preference representation is obtained;
[0012] Based on the user preference representation, predict the user's interaction score with candidate items and obtain recommendation results.
[0013] Optionally, obtaining the item embedding representation of each item based on the identification information of each item in the user interaction item set includes:
[0014] Based on the one-hot encoding of the items, the item embedding representation of each item is obtained by mapping through an embedding lookup layer.
[0015] Optionally, based on the input sequence, encoding processing is performed using a state-space model to obtain the encoded vector, including:
[0016] Based on the input sequence, the encoding vector is obtained by performing layer-by-layer transformation through several stacked sub-layers, wherein each sub-layer contains a selective state space update module and a position-level feedforward network module.
[0017] Optionally, obtaining the encoded vector by performing a layer-by-layer transformation through several stacked sub-layers based on the input sequence includes:
[0018] Based on the hidden state output by the (a-1)th sub-layer, the context-aware sequence is obtained through the selective state space update module;
[0019] Based on the context-aware sequence, the hidden state of the output of the a-th sub-layer is obtained through the position-level feedforward network module;
[0020] The transformation is performed sequentially from the first to the Nth sub-layer to obtain the encoding vector output by the Nth sub-layer.
[0021] Optionally, obtaining the context-aware sequence through the selective state space update module includes:
[0022] Based on the current input sequence, obtain the dynamically generated input selectivity matrix and output selectivity matrix;
[0023] Based on the input selectivity matrix, the globally shared state transition matrix, and the hidden state of the previous position, obtain the hidden state of the current position;
[0024] Based on the output selectivity matrix and the hidden state at the current position, obtain the output representation of the current position;
[0025] The context-aware sequence is obtained by concatenating the output representations at each position.
[0026] Optionally, concatenating the output representations at each location to obtain the context-aware sequence includes:
[0027] ;
[0028] in, Let be the selective state-space update function for layer a. Let be the hidden state of the input to the a-th sub-layer. This represents the output of layer a at position 0. This represents the output of layer a at position 1. This represents the output of layer a at position n.
[0029] The output of the sub-layer is: ;
[0030] in, Let be the hidden state output by the a-th sub-layer. Let be the position-level feedforward network function of layer a.
[0031] Optionally, obtaining the hidden state of the output of the a-th sublayer through the position-level feedforward network module includes:
[0032] Based on the context-aware sequence, an intermediate representation is obtained by transforming it through a first linear layer and a nonlinear activation function.
[0033] Based on the intermediate representation, the hidden state of the output of the a-th sub-layer is obtained by transforming it through the second linear layer.
[0034] Optionally, based on the user preference representation, predicting the user's interaction score with candidate items includes:
[0035] Based on the user preference representation, a decoder network is used to map and obtain the predicted interaction scores between the user and all candidate items:
[0036] ;
[0037] in, For interactive scoring, (*) represents the decoder network. Encode vectors for users.
[0038] This invention also provides a collaborative filtering recommendation system based on a state-space model, comprising: a data acquisition module, a model processing module, and a result recommendation module;
[0039] The data acquisition module is used to obtain a set of user-interacted items based on the user's item interaction data; obtain the item embedding representation and user embedding representation of each item based on the identification information of each item in the set of user-interacted items and the user's identification information; and construct an input sequence based on the item embedding representation and user embedding representation.
[0040] The model processing module is used to perform encoding processing on the input sequence through a state-space model to obtain an encoding vector;
[0041] The result recommendation module is used to obtain a user preference representation based on the encoding vector; predict the interaction score between the user and the candidate items based on the user preference representation, and obtain a recommendation result.
[0042] Compared with the prior art, the present invention has the following advantages and technical effects:
[0043] This invention collects complete user interaction data on items from e-commerce websites, streaming media services, and social media platforms, performs data preprocessing to generate input sequences, models these sequences using the Mamba state-space model to generate encoded vectors, and finally reconstructs and optimizes these vectors to predict user preferences. This invention solves the problems of existing technologies that fail to comprehensively consider the global interaction history between users and items, resulting in incomplete understanding of user preferences, fragmented interaction context, loss of sequence preference information, and inaccurate prediction of user preferences. Attached Figure Description
[0044] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:
[0045] Figure 1 This is a flowchart of a collaborative filtering recommendation method based on a state-space model according to an embodiment of the present invention;
[0046] Figure 2 This is a schematic diagram of a collaborative filtering recommendation system based on a state-space model according to an embodiment of the present invention. Detailed Implementation
[0047] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.
[0048] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.
[0049] This embodiment proposes a collaborative filtering recommendation method based on a state-space model, such as... Figure 1 As shown, the specific steps include:
[0050] Based on the user's item interaction data, obtain the set of items the user interacted with;
[0051] Based on the identification information of each item in the user interaction item set and the user's identification information, obtain the item embedding representation and user embedding representation of each item respectively;
[0052] Construct the input sequence based on the item embedding representation and the user embedding representation;
[0053] Based on the input sequence, the encoding vector is obtained by encoding through a state-space model.
[0054] Based on the encoded vector, obtain the user preference representation;
[0055] Based on user preference representations, predict the user's interaction score with candidate items to obtain recommendation results.
[0056] Specifically, for users First, aggregate all its interactive items into a complete set. , where is the number of items the user interacted with.
[0057] More specifically, each item After one-hot encoding, it is mapped to dense embedding through a lookup table layer. This facilitates subsequent neural network processing.
[0058] Mimicking the [CLS] token design in BERT, the user context is injected into the first token: the user embedding is also obtained through a table lookup layer. .
[0059] Furthermore, based on the item embedding representation and the user embedding representation, the input sequence is constructed as follows:
[0060] The user embedding is placed at the beginning of the sequence, and the item embeddings are concatenated sequentially to obtain the input sequence:
[0061] ;
[0062] Where x is the input sequence, Embedded representation for users, All are represented by embedded items. ( (for the embedded dimension) The length of the structure is sequence.
[0063] Furthermore, based on the input sequence, encoding is performed using a state-space model to obtain the encoded vector, which includes:
[0064] Based on the input sequence, the encoding vector is obtained by performing layer-by-layer transformation through several stacked sub-layers. Each sub-layer contains a selective state space update module and a position-level feedforward network module.
[0065] Furthermore, based on the input sequence, the encoding vector is obtained by performing layer-by-layer transformations through several stacked sub-layers, including:
[0066] Based on the hidden state output by the (a-1)th sub-layer, the context-aware sequence is obtained through the selective state space update module;
[0067] Based on the context-aware sequence, the hidden state of the output of the a-th sub-layer is obtained through the position-level feedforward network module;
[0068] The encoding vector of the Nth sub-layer is obtained by performing transformations sequentially from the 1st to the Nth sub-layer.
[0069] Specifically, MambaCF uses the Mamba state-space model as the encoder, which then processes the preprocessed sequence. Compressed into a fixed-length hidden vector It encapsulates user preferences based on the user's complete interaction history. Formally, the Mamba encoder is defined as a function... :
[0070] ;
[0071] In the formula, It is a fixed-length hidden vector used to summarize the user's complete interaction preferences.
[0072] The Mamba model consists of N stacked sublayers, each containing a Selective State Update Mechanism (SSM) and a Location-Level Feedforward Network (FFN). Defined as a sub-layer function, it consists of the following:
[0073] ;
[0074] In the formula, Indicates the first Sub-layer transformation, the last sub-layer The final output is .
[0075] Each sub-layer It consists of two main computational modules: Selective StateSpace Update (SSM) and Position-wise Feed-Forward Network (FFN).
[0076] Order No. The input to each sub-layer is the hidden state. (Initial conditions) = The output is The formal definition of the mapping of this sublayer is:
[0077] ;
[0078] In the formula, It is the first Selective state-space update of layers; It is the first FFN of the layer.
[0079] The selective state-space mechanism involves a hidden state based on the input sequence. For the input sequence (In the formula: each) (corresponding to) or The state update follows the discretized state-space equations as follows:
[0080] ;
[0081] ;
[0082] In the formula, Let be the hidden state of layer a at position t, with dimension . ; This represents the hidden state of layer a-1 at position t-1. It is a globally shared, trainable state transition matrix; The input-driven selectivity matrix assigns the model an input-dependent receptive field and gain. This represents the output of the SSM at position t.
[0083] All are from the current input Dynamic generation, typically calculated using the following linear projection:
[0084] ;
[0085] ;
[0086] In the formula, These are the projection weights of the input matrix; These are the projection weights of the output matrix.
[0087] The output of the entire sequence after SSM is the concatenation of the corresponding output vectors:
[0088] ;
[0089] Right now: The operator ultimately returns a selective context-aware representation sequence of the same length as the input, which can be directly used as input for subsequent FFNs.
[0090] Context-aware sequence of SSM output by feedforward network (FFN) After refinement, a typical FFN with two linear layers and one nonlinear layer is as follows:
[0091] ;
[0092] In the formula, , (Intermediate dimension) ); , ; It is a non-linear activation function.
[0093] Therefore, the sublayer output is defined as follows:
[0094] ;
[0095] After forward computation through N sub-layers, the final hidden state is... Output from the last sub-layer Generate. To obtain a user-level representation, take... The first token vector is used as the user encoding vector. ,as follows:
[0096] .
[0097] Furthermore, obtaining the context-aware sequence through the selective state-space update module includes:
[0098] Based on the current input sequence, obtain the dynamically generated input selectivity matrix and output selectivity matrix;
[0099] Based on the input selectivity matrix, the globally shared state transition matrix, and the hidden state of the previous position, obtain the hidden state of the current position;
[0100] Based on the output selectivity matrix and the hidden state at the current position, obtain the output representation of the current position;
[0101] The context-aware sequence is obtained by concatenating the output representations at each position.
[0102] Furthermore, obtaining the hidden state of the output of the a-th sublayer through the position-level feedforward network module includes:
[0103] Based on the context-aware sequence, an intermediate representation is obtained by transforming it through a first linear layer and a non-linear activation function.
[0104] Based on the intermediate representation, the hidden state of the output of the a-th sub-layer is obtained by transforming it through the second linear layer.
[0105] Furthermore, based on user preference representations, the predicted user interaction score with candidate items includes:
[0106] Based on user preference representations, a decoder network is used to map and obtain predicted interaction scores between the user and all candidate items.
[0107] Specifically, to ensure the encoding vector This approach preserves the core features of the original input and introduces a reconstruction objective. The specific steps are as follows:
[0108] User-coded vectors Input decoder network ;
[0109] Furthermore, the decoder is composed of multiple layers of fully connected (FC) stacked together, responsible for processing the compressed representation. Map back to the original input dimension;
[0110] For any user The decoder is then connected to an output layer to calculate interaction scores for all items. The definition is as follows:
[0111] ;
[0112] The loss is predicted using binary cross-entropy (BCE). In users Candidate product set Above calculation:
[0113] ;
[0114] In the formula, For real interaction tags: If the user With items If the interaction is observed, the value is 1; otherwise, it is 0.
[0115] The following is a detailed description of this embodiment with reference to the accompanying drawings:
[0116] The method of this embodiment includes the following steps:
[0117] Collect complete item interaction data from users across e-commerce websites, streaming services, and social media platforms. First, aggregate all its interactive items into a complete set. ,in The number of items interacted with by this user. Each item After one-hot encoding, it is mapped to dense embedding through a lookup table layer. This method reduces dimensionality while preserving semantic information, making it easier for subsequent neural network processing.
[0118] Furthermore, mimicking the [CLS] token design in BERT, the user context is injected into the first token: the user embedding is also obtained through a table lookup layer. The final Mamba encoder's original input sequence is generated by concatenating it using the following formula:
[0119] ;
[0120] In the formula, ( (for the embedded dimension) The length of the structure is The input sequence.
[0121] Unlike sequence-based recommendations that emphasize temporal order, this embodiment treats the item set as a whole, focusing on the collective impact of all interactions rather than their sequential order. This allows for a more effective capture of the inherent patterns in user-item interactions, leveraging aggregated context to improve prediction accuracy. By jointly incorporating user and item embeddings, the model focuses on accurate correlations within the data, avoiding interference from potential temporal biases in sequence processing.
[0122] Furthermore, MambaCF uses the Mamba state-space model as an encoder for the input sequence. Mamba enables efficient modeling by maintaining a latent state vector that evolves with the input, capturing long-range dependencies and complex relationships between items in linear complexity.
[0123] The encoder will preprocess the sequence Compressed into a fixed-length hidden vector It encapsulates user preferences based on the user's complete item interaction history. Formally, the Mamba encoder is defined as a function... :
[0124] ;
[0125] In the formula, It is a fixed-length hidden vector used to summarize the user's complete interaction preferences.
[0126] The Mamba model consists of N stacked sublayers, each containing a Selective State Update Mechanism (SSM) and a Location-Level Feedforward Network (FFN). Defined as a sub-layer function, it consists of the following:
[0127] ;
[0128] In the formula, Indicates the first Sub-layer transformation, the last sub-layer The final output is .
[0129] Furthermore, each sub-layer It consists of two main computational modules: Selective State Space Update (SSM) and Position-wise Feed-Forward Network (FFN).
[0130] Order No. The input to each sub-layer is the hidden state. (Initial conditions) = The output is The formal definition of the mapping of this sublayer is:
[0131] ;
[0132] In the formula, It is the first Selective state-space update of layers; It is the first Location-level feedforward network of the layer.
[0133] Furthermore, the selective state-space mechanism involves a hidden state based on the input sequence. For the input sequence (In the formula: each) (corresponding to) or The state update follows the discretized state-space equations as follows:
[0134] ;
[0135] ;
[0136] In the formula, Let be the hidden state of layer a at position t, with dimension . ; It is a globally shared, trainable state transition matrix; The input-driven selectivity matrix assigns the model an input-dependent receptive field and gain. This represents the output of the SSM at position t.
[0137] All are from the current input Dynamic generation, typically calculated using the following linear projection:
[0138] ;
[0139] ;
[0140] In the formula, These are the projection weights of the input matrix; These are the projection weights of the output matrix.
[0141] The output of the entire sequence after SSM is the concatenation of the corresponding output vectors:
[0142] ;
[0143] Right now: The operator ultimately returns a selective context-aware representation sequence of the same length as the input, which can be directly used as input for subsequent FFNs.
[0144] Context-aware sequence of SSM output by feedforward network (FFN) After refinement, a typical FFN with two linear layers and one nonlinear layer is as follows:
[0145] ;
[0146] In the formula, , (Intermediate dimension) ); , ; It is a non-linear activation function: .
[0147] Therefore, the sublayer output is defined as follows:
[0148] ;
[0149] After forward computation through N sub-layers, the final hidden state is... Output from the last sub-layer Generate. To obtain a user-level representation, take... The first token vector is used as the user preference vector, as follows:
[0150] ;
[0151] The selective update mechanism dynamically adjusts the latent state based on the relevance of each input element, while FFN refines the representation. The encoding process utilizes Mamba... The ability to model global interactions between all projects in the system, through user embedding Weighting is applied. By treating the input as a unified set rather than a temporally ordered sequence, the encoder learns a representation that reflects the collective influence of user preferences, making it well-suited for CF tasks.
[0152] Furthermore, to ensure the encoding vector This approach preserves the core features of the original input and introduces a reconstruction objective. The specific steps are as follows:
[0153] User-coded vectors Input decoder network ;
[0154] Furthermore, the decoder is composed of multiple layers of fully connected (FC) stacked together, responsible for processing the compressed representation. Map back to the original input dimension;
[0155] For any user The decoder is then connected to an output layer to calculate interaction scores for all items. The definition is as follows:
[0156] ;
[0157] The prediction loss uses binary cross-entropy (BCE) on the user. Candidate product set Above calculation:
[0158] ;
[0159] In the formula, For real interaction tags: If the user With items If the interaction is observed, the value is 1; otherwise, it is 0.
[0160] Table 1 shows the performance comparison of MambaCF (Mamba-based Collaborative Filtering State-Space Model) with other models on different datasets after experiments using the MambaCF model from this embodiment. In this embodiment, the experimental results demonstrate the overall performance comparison of MambaCF and baseline methods on three datasets (Lastfm, ML-100K, and A-Baby). The evaluation is based on two key metrics: HR@10 and NDCG@10, which measure the ability to retrieve relevant items from the top 10 recommendations and effectively rank them. The results are summarized in Table 1. MambaCF consistently outperforms all baseline methods on all evaluated datasets, demonstrating its superior capabilities in collaborative filtering tasks.
[0161] Table 1
[0162]
[0163] MambaCF leverages the Mamba state-space model to dynamically adjust its latent states based on the input sequence, capturing long-range dependencies and global interactions between items. MambaCF's holistic encoding of the entire set of interactions provides a more comprehensive view of user preferences, improving retrieval accuracy (HR@10) and ranking accuracy (NDCG@10). These results demonstrate the significant potential of state-space models like MambaCF for improving recommender systems, warranting further exploration of their adaptability and optimization for various recommender scenarios.
[0164] This embodiment also provides a collaborative filtering recommendation system based on a state-space model, including: a data acquisition module, a model processing module, and a result recommendation module;
[0165] The data acquisition module is used to obtain a set of user-interacted items based on the user's item interaction data; to obtain the item embedding representation and user embedding representation of each item based on the identification information of each item in the set of user-interacted items and the user's identification information; and to construct the input sequence based on the item embedding representation and user embedding representation.
[0166] The model processing module is used to perform encoding processing on the input sequence through the state space model to obtain the encoded vector;
[0167] The result recommendation module is used to obtain user preference representations based on the encoded vectors; based on the user preference representations, it predicts the interaction scores between users and candidate items and obtains recommendation results.
[0168] like Figure 2 As shown, the platform includes a data acquisition layer, a data processing layer, a model inference layer, a business logic layer, and a user interface layer.
[0169] In the data acquisition layer, the collected data is divided into user behavior data, product static data, and user static data. User behavior data can include user actions such as clicks, adding to cart, placing orders, and adding to favorites. Product static data can include basic information such as product ID, category, brand, and price. User static data can include basic information such as user ID, gender, age, and city tier.
[0170] Furthermore, in the data processing layer, the collected data is processed, and user data is combined with item data to generate the original input sequence.
[0171] Furthermore, in the model inference layer, the processed input sequence is input into the MambaCF encoder to generate an encoding vector, and finally, through the output layer, recommended products are generated.
[0172] Furthermore, at the business logic layer, recommendation results can be adjusted and rearranged according to business strategies to ensure diversity and rationality. Simultaneously, user feedback data on recommendation results can be collected for online learning and optimization of the model.
[0173] Furthermore, in the user interface layer, the recommendation results are displayed to the user, providing a personalized list of product recommendations. At the same time, users can perform actions such as clicking on the recommendations, adding them to their cart, and placing orders. This behavioral data will be collected in real time and fed back to the system.
[0174] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.
Claims
1. A collaborative filtering recommendation method based on a state-space model, characterized in that, include: Based on the user's item interaction data, obtain the set of items the user interacted with; Based on the identification information of each item in the user interaction item set and the user's identification information, obtain the item embedding representation and user embedding representation of each item respectively; Construct the input sequence based on the item embedding representation and the user embedding representation; Based on the input sequence, an encoding vector is obtained by encoding through a state-space model. Based on the input sequence, encoding is performed using a state-space model to obtain the encoded vector, including: Based on the input sequence, the encoding vector is obtained by performing layer-by-layer transformation through several stacked sub-layers, wherein each sub-layer contains a selective state space update module and a position-level feedforward network module; Based on the input sequence, the encoded vector is obtained by performing a layer-by-layer transformation through several stacked sub-layers, including: Based on the hidden state output by the (a-1)th sub-layer, the context-aware sequence is obtained through the selective state space update module; Based on the context-aware sequence, the hidden state of the output of the a-th sub-layer is obtained through the position-level feedforward network module; The transformation is performed sequentially from the 1st to the Nth sub-layer to obtain the encoding vector output by the Nth sub-layer; Based on the encoded vector, the user preference representation is obtained; Based on the user preference representation, predict the user's interaction score with candidate items and obtain recommendation results; Obtaining the context-aware sequence through the selective state space update module includes: Based on the current input sequence, obtain the dynamically generated input selectivity matrix and output selectivity matrix; Based on the input selectivity matrix, the globally shared state transition matrix, and the hidden state of the previous position, obtain the hidden state of the current position; Based on the output selectivity matrix and the hidden state at the current position, obtain the output representation of the current position; The context-aware sequence is obtained by concatenating the output representations of each position. The context-aware sequence is obtained by concatenating the output representations at each position. ； in, Let be the selective state-space update function for layer a. Let be the hidden state of the input to the a-th sub-layer. This represents the output of layer a at position 0. This represents the output of layer a at position 1. This represents the output of layer a at position n. The output of the sub-layer is: ; in, Let be the hidden state output by the a-th sub-layer. Let be the position-level feedforward network function of layer a; Obtaining the hidden state of the output of the a-th sublayer through the position-level feedforward network module includes: Based on the context-aware sequence, an intermediate representation is obtained by transforming it through a first linear layer and a nonlinear activation function. Based on the intermediate representation, the hidden state of the output of the a-th sub-layer is obtained by transforming it through the second linear layer; Based on the user preference representation, the predicted user interaction score with candidate items includes: Based on the user preference representation, a decoder network is used to map and obtain the predicted interaction scores between the user and all candidate items: ； in, For interactive scoring, ( ) is the decoder network. Encode vectors for users.
2. The collaborative filtering recommendation method based on a state-space model according to claim 1, characterized in that, Based on the identifier information of each item in the user interaction item set, the item embedding representation of each item is obtained, including: Based on the one-hot encoding of the items, the item embedding representation of each item is obtained by mapping through an embedding lookup layer.
3. A collaborative filtering recommendation system based on a state-space model implemented according to the method of any one of claims 1-2, characterized in that, include: Data acquisition module, model processing module, and result recommendation module; The data acquisition module is used to obtain a set of user-interacted items based on the user's item interaction data; obtain the item embedding representation and user embedding representation of each item based on the identification information of each item in the set of user-interacted items and the user's identification information; and construct an input sequence based on the item embedding representation and user embedding representation. The model processing module is used to perform encoding processing on the input sequence through a state-space model to obtain an encoding vector; The result recommendation module is used to obtain a user preference representation based on the encoding vector; predict the interaction score between the user and the candidate items based on the user preference representation, and obtain a recommendation result.