A cross-domain sequential recommendation method based on causal inference and preference evolution

Abstract

This invention belongs to the field of cross-domain sequence recommendation technology, and discloses a cross-domain sequence recommendation method based on causal inference and preference evolution. It enhances the semantic information of items using a large language model by employing prompt word templates designed based on cross-domain co-occurrence frequency. Intra-domain preferences are obtained through domain-specific interest evolution model for intra-domain sequence preference learning; cross-domain preferences are obtained through a temporal-domain dual-conditional expert mechanism for cross-domain sequence preference learning. A causal debiasing design decouples user intra-domain preferences, user cross-domain preferences, and user intra-domain activity and cross-domain activity, and uses this debiasing to generate recommendations, improving the accuracy of cross-domain recommendation results. This method effectively overcomes the limitations of existing methods by leveraging a causal-enhanced preference learning mechanism, achieving state-of-the-art performance in both single-domain and cross-domain sequence recommendation benchmarks.

Description

Technical Field

[0001] This invention relates to the field of cross-domain sequence recommendation technology, and in particular to a cross-domain sequence recommendation method based on causal inference and preference evolution. Background Technology

[0002] Sequence recommendation, aiming to predict a user's next potentially interesting item based on their historical interaction sequences, has become an important research direction in the field of recommender systems in recent years. However, a key bottleneck in practical applications lies in the problem of data sparsity: most users have a limited number of interactions within a single domain, leading to overfitting of traditional sequence recommendation models and limiting their generalization performance. To address this challenge, cross-domain sequence recommendation has emerged. By introducing auxiliary interaction information from multiple related domains, it effectively alleviates the data sparsity problem in the target domain, becoming a solution with broad application prospects.

[0003] Unlike traditional sequence recommendation, which models only a single domain, cross-domain sequence recommendation can dynamically capture and integrate user behavior patterns across different domains, thereby generating recommendation results that better match the user's true intentions and context. For example, user interactions in the clothing domain (such as browsing shirts and sneakers sequentially) and the electronics domain (such as showing interest in smartwatches) may be difficult to establish effective connections if analyzed independently within each domain; however, with the help of a cross-domain sequence recommendation model, it is possible to identify the common preferences implied behind these behaviors, such as "preferring an active lifestyle" or "favoring high-performance technological equipment," thereby uncovering deeper user interests that are difficult to capture by single-domain models.

[0004] Existing cross-domain sequence recommendation methods can be broadly divided into two mainstream paradigms based on their core technologies: one is the classic method that mainly relies on collaborative filtering signals, and the other is the method that utilizes the capabilities of large language models.

[0005] The first type of method builds upon traditional recommendation techniques. Its core idea is to align and transfer collaborative signals between different domains through various machine learning techniques. Chinese patent CN120670651A proposes a cold-start recommendation method based on dynamic intent perception and cross-domain knowledge transfer. This method decomposes the representations of users and items into multiple intent subspaces by decoupling the feature extractor through graph convolutional neural networks, achieving fine-grained intent alignment. Chinese patent CN119415770A proposes a cold-start item recommendation method based on triple contrastive learning. This method optimizes the intra-domain and cross-domain preference representations of users by constructing self-contrast learning tasks, source-domain contrastive learning tasks, and cross-domain contrastive learning tasks. As large language models demonstrate powerful capabilities in general knowledge and contextual reasoning, the second type of emerging method begins to explore its application potential in this field. These studies mainly proceed in two directions: one is to directly utilize the native capabilities of large language models for end-to-end recommendation generation, such as the paper "Uncovering Cross-Domain Recommendation Ability of Large Language Models" (WWW2025), which uses refined prompt template design to enable large models to transfer cross-domain information and generate recommendation results. Secondly, large language models are used as enhancement components to assist and improve the performance of traditional recommendation models through semantic reasoning, thereby achieving a deep understanding of users' cross-domain intentions. For example, Chinese patent CN118898269B proposes an adaptive knowledge-enhanced multi-domain recommendation method based on large language models. This method calls upon large language models to learn features based on domain-shared and domain-specific prompt templates, obtaining the target user's domain-invariant features and domain-specific preference features, effectively improving the predictive performance of multi-domain recommendation models.

[0006] While these methods demonstrate good performance in cross-domain sequence recommendation, they face three main heterogeneity challenges in practical applications: semantics, temporal dynamics, and user activity. First, existing methods fail to bridge the cross-domain semantic gap. For example, in fitness scenarios, when users frequently interact with Disney-branded sneakers (clothing domain) and Disney-themed smartwatches (electronics domain), existing models cannot infer the shared implicit semantic attributes (such as the combination of "Disney IP" and "professional fitness"), remaining only at the surface category level. Second, existing methods cannot uniformly capture cross-domain preferences evolving across different domains. For instance, user preferences in the clothing domain may iterate rapidly, shifting from "rock" to "sports" and then to "commuter" within a year, exhibiting short-term fluctuations. In the electronics domain, although users may continuously purchase different categories such as smartwatches, high-performance computers, and professional cameras, their underlying core needs (such as "pursuing high-performance technological equipment") tend to remain stable over longer timescales. Current methods cannot adaptively capture the heterogeneous evolutionary rhythms across different domains. Finally, models are often affected by biases in user activity levels. A typical scenario involves users interacting frequently in the apparel domain (e.g., browsing or purchasing sports equipment weekly) but sparsely in the electronics domain (e.g., checking their smartwatch only once every six months). If the model simply relies on interaction frequency as a proxy for the strength of interest, it may overemphasize apparel preferences while ignoring the user's true underlying interests in the electronics domain—even though both may share the same cross-domain motivation of a "sports lifestyle." Current methods cannot decouple the user's true interest migration path from activity differences, thus resulting in recommendation bias due to uneven behavior frequency. Therefore, effectively bridging these differences is crucial for achieving high-quality knowledge transfer. Summary of the Invention

[0007] The purpose of this invention is to provide a cross-domain sequence recommendation method based on causal inference and preference evolution. Specifically, it enhances the semantic information of items by designing cue word templates based on cross-domain co-occurrence frequency and utilizing a large language model. Furthermore, it employs a temporal-domain dual-conditional hybrid expert mechanism to effectively capture users' rapidly evolving and slowly changing preference patterns through a unified representation. Finally, it designs a causal-enhanced preference learning strategy to effectively eliminate biases caused by activity levels by decoupling spurious causal relationships from genuine interests, thereby ensuring the robustness of cross-domain recommendation.

[0008] The technical solution of this invention is as follows: A cross-domain sequence recommendation method based on causal inference and preference evolution, comprising the following steps:

[0009] Step 1: Obtain user intra-domain interaction sequences and cross-domain interaction sequences, and calculate the cross-domain co-occurrence frequency;

[0010] Step 2: Design prompt word templates based on cross-domain co-occurrence frequency, and enhance the semantic information of items using a large language model to obtain the semantic embedding vector of items;

[0011] Step 3: Based on the semantic embedding vectors of items, intra-domain preferences are obtained by learning intra-domain sequence preferences through a domain-specific interest evolution model; cross-domain preferences are obtained by learning cross-domain sequence preferences through a temporal-domain dual-conditional expert mechanism.

[0012] Step 4: For both intra-domain and cross-domain interaction sequences, use a fixed time window to count the total number of interactions within each time window. Using the global maximum value among all time windows as the standard, normalize the number of intra-domain and cross-domain interactions respectively. The normalized results are used as the user's dynamically changing intra-domain activity and cross-domain activity.

[0013] Step 5: Perform causal bias removal based on intra-domain preferences, cross-domain preferences, intra-domain activity, and cross-domain activity to generate recommendations.

[0014] The cross-domain co-occurrence frequency is calculated as follows: cross-domain co-occurrence relationships are mined using a large language model to obtain the functional associations between items; when item A and item B appear together in any user sequence, the corresponding cross-domain co-occurrence frequency increases.

[0015] The prompt word template specifically involves: requiring the large language model to generate a description by combining the item's own information with cross-domain co-occurrence relationships; for high-frequency item co-occurrence pairs with a cross-domain co-occurrence frequency greater than a set threshold, instructing the large language model to explore their shared usage scenarios and complementary attributes; for low-frequency item co-occurrence pairs with a cross-domain co-occurrence frequency not exceeding a set threshold, instructing the large language model to analyze their potential contextual associations; and finally, the large language model integrates all the above information to generate enhanced semantic information for the item, not exceeding the maximum word limit, and generates an item semantic embedding vector based on this enhanced semantic information. .

[0016] The item's own information includes the item name and the item's original category; the cross-domain co-occurrence relationship is a list of the top 5 items with the highest cross-domain co-occurrence frequency.

[0017] The domain interaction sequence consists of two parts: first, the domain behavior sequence of user interaction with items in the domain; and second, the domain time sequence consisting of the time of each interaction.

[0018] The domain-specific interest evolution model extracts and models domain-specific behavior sequences and domain-specific time sequences from domain-specific interaction sequences, thereby learning dynamic user domain-specific preferences.

[0019] Specifically, this refers to: semantic embedding vectors of items Principal component analysis is performed for dimensionality reduction to extract low-dimensional item embedding vectors that retain their dominant semantic features. Each item in the user's in-domain behavior sequence is represented by its corresponding low-dimensional item embedding vector. An encoder based on the SASRec architecture is used to model the user's in-domain behavior sequence through a self-attention mechanism, generating in-domain behavior embeddings. The formula for the self-attention mechanism is as follows:

[0020]

[0021] Where Q, K, and V represent the query embedding representation, key embedding representation, and value embedding representation, respectively; d represents the embedding dimension. This is a self-attention mechanism; For activation functions;

[0022] Based on the intra-domain time series in the intra-domain interaction sequence First, calculate the relative time interval in the X domain: By processing relative time intervals through time-series coding, corresponding intra-domain time embeddings are generated. The timing coding formula is as follows:

[0023]

[0024] Set the hyperparameter to 1000; This is the time transformation matrix;

[0025] Embedding behavior within the splicing domain Intradomain time embedding Obtain the in-domain preference representation For another domain Y, its in-domain preference representation It is obtained through exactly the same steps.

[0026] The cross-domain interaction sequence consists of two parts: one is the cross-domain behavior sequence of user interaction with cross-domain items, and the other is the cross-domain time sequence composed of the time of each interaction. Based on the time-domain dual-condition hybrid expert mechanism, the cross-domain behavior sequence and the cross-domain time sequence are extracted and modeled from the cross-domain interaction sequence, and then the dynamic user cross-domain preferences are learned.

[0027] Specifically, firstly, the semantic embedding vector of the item is... After linear transformation to adapt to the recommendation task, the formula is as follows:

[0028]

[0029] in, Represents the semantic embedding vector of an item. The Middle Embedding vectors of each item Indicates the first After the above linear transformation, each item is used as a global embedding representation for the recommendation task. and These are the weight matrices that change linearly at different levels; and These are the bias vectors for different levels of linear transformation;

[0030] A self-attention mechanism is used to model cross-domain behavior sequences to obtain cross-domain behavior embeddings. The original cross-domain time embedding is obtained by using time-series coding on the cross-domain time series. ;

[0031] For each absolute timestamp in the cross-domain time series, an absolute time embedding is learned. A domain embedding is generated using a binary domain identifier, with 0 representing the source domain and 1 representing the target domain. The domain embedding and the absolute time embedding are fused as the gating weights of a time-domain dual-conditional hybrid expert mechanism. Through weighted fusion by experts, a cross-domain time embedding adjusted on a unified time scale is finally output. Cross-domain behavioral embeddings are then incorporated. and adjusted cross-domain time embedding The cross-domain preference representation is obtained by splicing. .

[0032] The adjusted cross-domain time embedding acquisition process is as follows:

[0033] Each absolute timestamp in a cross-domain time series is decomposed into year, month, and date. A unified time embedding is constructed through a three-layer embedding structure of year embedding, month embedding, and date embedding.

[0034]

[0035] in, , , These represent the year embedding vector, month embedding vector, and date embedding vector, respectively. and These are the weight matrices for different levels of linear transformation; and These are the bias vectors for different levels of linear transformation; For computed output, a uniform temporal embedding;

[0036] Adopting binary domain identifiers As a semantic anchor, Represents the source domain. Represent the target domain, and assign each binary identifier Mapping to Domain Embedding Attention mechanism Embedding the domain in a linear transformation manner and absolute time embedding Fusion; The specific calculation method for gating weights is as follows:

[0037]

[0038] in, and These are the weight matrix and the bias vector, respectively. Indicates the gating weight;

[0039] The cross-domain temporal information representation adjusted by the computational time-domain biconditional hybrid expert mechanism is obtained by weighted fusion of the cross-domain temporal embedding representations of multiple experts, as shown in the following formula:

[0040]

[0041] in, It is the first The weight parameters of an expert network, yes Assigned to the The gating weight of each expert, yes The Middle Cross-domain time embedding representation for each user It is the first A cross-domain temporal embedding representation of a user after adjustment by a time-domain biconditional expert mechanism.

[0042] The causal debiasing method is a causal-enhanced preference learning approach. It decomposes user-domain-specific and user-cross-domain preferences using a backdoor adjustment method, decoupling them from user-domain-specific and cross-domain activity. In cross-domain sequence recommendation scenarios, two confounding factors exist: user-domain-specific activity... and user cross-domain activity ;

[0043] The confounding factor introduces spurious correlations through two biased paths: one is the activity biased path. , The first is the intra-domain preference representation, where the activity-based bias path improves the recommendation accuracy for highly active users; the second is the cross-domain top user bias path. , The cross-domain preference representation causes the recommendation results to be overly biased towards users with frequent interactions; a backdoor adjustment strategy is introduced to correct this bias, and the causal effect of the correction is estimated as follows:

[0044]

[0045] In this process (a), based on Bayes' theorem, the intervention conditional probability is transformed into a summation of conditional probabilities based on observation data. Indicates to and Causal graph model after intervention; Process (b) blocking backdoor pathway: for and Intervention, i.e. and Cross-domain user activity and user domain activity Cross-domain preference representation Intra-domain preference representation The conditional dependency between them is severed, and their conditional probability degenerates into prior probability: Process (c) based on causal mechanism invariance: from post-intervention preference representation and confounding factors to recommendation outcome The causal mechanism remains unchanged; process (d) calculates the causal effect: Measuring preferences indicates the causal effect on recommendation outcomes. and The mathematical expectation representing the confounding factor;

[0046] Ultimately, with This represents the final recommendation result in the X domain, where... Linear fusion is employed; the binary cross-entropy loss function for the X domain is defined as follows:

[0047]

[0048] in, Indicates the predicted recommendation result Chinese users The probability of interacting with an item in domain X. Indicates user Real-world interaction tags with items in the X domain. Let X be the loss function for the domain X. By minimizing this loss function, the probability of predicting user behavior in the domain X is obtained. As close as possible to real interactive labels ;

[0049] The loss function is obtained in the same way in the Y domain. Then the cross-domain sequence recommendation loss function is The accuracy of cross-domain recommendation results can be improved by optimizing the cross-domain sequence recommendation loss function.

[0050] The beneficial effects of this invention are as follows: This invention proposes a cross-domain sequence recommendation method based on causal inference and preference evolution. By using prompt word templates designed based on cross-domain co-occurrence frequency, a large language model is used to enhance the semantic information of items. Dynamic user intra-domain preferences and cross-domain preferences are learned through a domain-specific interest evolution model and a temporal-domain biconditional hybrid expert mechanism. Causal debiasing is designed to decouple user intra-domain preferences, user cross-domain preferences, and user intra-domain activity and cross-domain activity, and this debiasing is used to generate recommendations, improving the accuracy of cross-domain recommendation results. Attached Figure Description

[0051] Figure 1 This is a schematic diagram illustrating the recommendation performance of the Amazon Movies-Books dataset, with ExcelRec representing the method of this invention.

[0052] Figure 2 This is a schematic diagram illustrating the recommendation performance of the Amazon Apparel-Appliances domain dataset.

[0053] Figure 3 This is a cause-and-effect graph model diagram;

[0054] Figure 4 This is a flowchart of a cross-domain sequence recommendation method based on causal inference and preference evolution. Detailed Implementation

[0055] The core of this invention lies in constructing a cross-domain sequence recommendation method based on causal inference and preference evolution. Taking the movie domain and book domain as an example, firstly, the cross-domain co-occurrence frequency of all items in the movie domain and book domain is statistically analyzed. Based on the cross-domain co-occurrence frequency, prompt word templates are designed, and semantic information of all items is enhanced using a large language model. From the user's intra-domain interaction sequence in the movie domain, intra-domain behavior sequences and intra-domain time sequences are extracted and modeled respectively. Intra-domain preferences in the movie domain are obtained through a domain-specific interest evolution model. Intra-domain preferences in the book domain can be obtained in the same way. From the user's cross-domain interaction sequence in the movie domain and book domain, cross-domain behavior sequences and cross-domain time sequences are extracted and modeled respectively. Through a time-domain dual-condition hybrid expert mechanism, the user's cross-domain preferences in the movie domain and book domain are obtained. This method calculates user activity levels within the movie domain, book domain, and cross-domain activity (movie domain - book domain) separately. By employing causal bias removal, it decouples movie domain preferences, book domain preferences, and cross-domain preferences from movie domain activity levels, book domain activity levels, and cross-domain activity (movie domain - book domain), generating recommendation results for both movie and book domains. The following technical means overcome the limitations of existing cross-domain sequence recommendation methods:

[0056] Obtain user intra-domain interaction sequences and cross-domain interaction sequences, and calculate cross-domain co-occurrence frequency;

[0057] A large language model-enhanced co-occurrence relation mining tool models the functional associations between items. This method is implemented in two phases: cross-domain co-occurrence frequency calculation—when items A and B co-occur in any user sequence, the corresponding co-occurrence count increases;

[0058] Based on cross-domain co-occurrence frequency, we designed prompt word templates and enhanced the semantic information of items using a large language model to obtain item semantic embedding vectors. The description generation enhanced by the large language model prioritizes the top five high-frequency co-occurring related objects for each item through carefully designed prompt word templates. To ensure description quality, we implemented a differentiated strategy: for high-frequency item co-occurrence pairs, we focused on mining shared attributes driving user interaction; while for low-frequency item co-occurrence pairs, we strengthened the semantic construction of their contextual associations. After obtaining the enhanced semantic information of the items, we used a large language model to obtain the item semantic embedding vectors. .

[0059] Based on the semantic embedding vectors of items, intra-domain preferences are obtained through intra-domain sequence preference learning using a domain-specific interest evolution model; cross-domain preferences are obtained through cross-domain sequence preference learning using a temporal-domain dual-conditional expert mechanism.

[0060] The domain interaction sequence consists of two parts: first, the domain behavior sequence of user interaction with items in the domain; and second, the domain time sequence consisting of the time of each interaction.

[0061] The domain-specific interest evolution model extracts and models domain-specific behavior sequences and domain-specific time sequences from domain-specific interaction sequences, thereby learning dynamic user domain-specific preferences.

[0062] Specifically, this refers to: semantic embedding vectors of items Principal component analysis is performed for dimensionality reduction to extract low-dimensional item embedding vectors that retain their dominant semantic features. Each item in the user's in-domain behavior sequence is represented by its corresponding low-dimensional item embedding vector. An encoder based on the SASRec architecture is used to model the user's in-domain behavior sequence through a self-attention mechanism, generating in-domain behavior embeddings. The formula for the self-attention mechanism is as follows:

[0063]

[0064] Where Q, K, and V represent the query embedding representation, key embedding representation, and value embedding representation, respectively; d represents the embedding dimension. This is a self-attention mechanism; For activation functions;

[0065] Based on the intra-domain time series in the intra-domain interaction sequence First, calculate the relative time interval in the X domain: By processing relative time intervals through time-series coding, corresponding intra-domain time embeddings are generated. The timing coding formula is as follows:

[0066]

[0067] Set the hyperparameter to 1000; This is the time transformation matrix;

[0068] Embedding behavior within the splicing domain Intradomain time embedding Obtain the in-domain preference representation For another domain Y, its in-domain preference representation It is obtained through exactly the same steps.

[0069] The cross-domain interaction sequence consists of two parts: one is the cross-domain behavior sequence of user interaction with cross-domain items, and the other is the cross-domain time sequence composed of the time of each interaction. Based on the time-domain dual-condition hybrid expert mechanism, the cross-domain behavior sequence and the cross-domain time sequence are extracted and modeled from the cross-domain interaction sequence, and then the dynamic user cross-domain preferences are learned.

[0070] Specifically, firstly, the semantic embedding vector of the item is... After linear transformation to adapt to the recommendation task, the formula is as follows:

[0071]

[0072] in, Represents the semantic embedding vector of an item. The Middle Embedding vectors of each item Indicates the first After the above linear transformation, each item is used as a global embedding representation for the recommendation task. and These are the weight matrices that change linearly at different levels; and These are the bias vectors for different levels of linear transformation;

[0073] A self-attention mechanism is used to model cross-domain behavior sequences to obtain cross-domain behavior embeddings. The original cross-domain time embedding is obtained by using time-series coding on the cross-domain time series. ;

[0074] For each absolute timestamp in the cross-domain time series, an absolute time embedding is learned. A domain embedding is generated using a binary domain identifier, with 0 representing the source domain and 1 representing the target domain. The domain embedding and the absolute time embedding are fused as the gating weights of a time-domain dual-conditional hybrid expert mechanism. Through weighted fusion by experts, a cross-domain time embedding adjusted on a unified time scale is finally output. Cross-domain behavioral embeddings are then incorporated. and adjusted cross-domain time embedding The cross-domain preference representation is obtained by splicing. .

[0075] The intra-domain interaction sequence and cross-domain interaction sequence are each divided into fixed time windows. The total number of interactions within each time window is counted. The global maximum value among all time windows is used as the standard to normalize the number of intra-domain interactions and cross-domain interactions. The normalized results are used as the user's dynamically changing intra-domain activity and cross-domain activity.

[0076] Based on intra-domain preferences, cross-domain preferences, intra-domain activity, and cross-domain activity, causal bias is removed to generate recommendations.

[0077] Furthermore, the adjusted cross-domain time embedding acquisition process is as follows:

[0078] Each absolute timestamp in a cross-domain time series is decomposed into year, month, and date. A unified time embedding is constructed through a three-layer embedding structure of year embedding, month embedding, and date embedding.

[0079]

[0080] in, , , These represent the year embedding vector, month embedding vector, and date embedding vector, respectively. and These are the weight matrices for different levels of linear transformation; and These are the bias vectors for different levels of linear transformation; For computed output, a uniform temporal embedding;

[0081] Adopting binary domain identifiers As a semantic anchor, Represents the source domain. Represent the target domain, and assign each binary identifier Mapping to Domain Embedding Attention mechanism Embedding the domain in a linear transformation manner and absolute time embedding Fusion; The specific calculation method for gating weights is as follows:

[0082]

[0083] in, and These are the weight matrix and the bias vector, respectively. Indicates the gating weight;

[0084] The cross-domain temporal information representation adjusted by the computational time-domain biconditional hybrid expert mechanism is obtained by weighted fusion of the cross-domain temporal embedding representations of multiple experts, as shown in the following formula:

[0085]

[0086] in, It is the first The weight parameters of an expert network, yes Assigned to the The gating weight of each expert, yes The Middle Cross-domain time embedding representation for each user It is the first A cross-domain temporal embedding representation of a user after adjustment by a time-domain biconditional expert mechanism.

[0087] Furthermore, the causal debiasing is a causal-enhanced preference learning method that decomposes user-domain in-domain preferences and user-cross-domain preferences through a backdoor adjustment method, decoupling them from user-domain in-domain activity and cross-domain activity. In cross-domain sequence recommendation scenarios, there are two confounding factors: user-domain in-domain activity... and user cross-domain activity ;like Figure 3 As shown, Figure 3 (a) The original causal graph G is obtained by causal bias removal. Figure 3 (b) The causal graph G' that has been interfered with.

[0088] The confounding factor introduces spurious correlations through two biased paths: one is the activity biased path. , The first is the intra-domain preference representation, where the activity-based bias path improves the recommendation accuracy for highly active users; the second is the cross-domain top user bias path. , The cross-domain preference representation causes the recommendation results to be overly biased towards users with frequent interactions; a backdoor adjustment strategy is introduced to correct this bias, and the causal effect of the correction is estimated as follows:

[0089]

[0090] In this process (a), based on Bayes' theorem, the intervention conditional probability is transformed into a summation of conditional probabilities based on observation data. Indicates to and Causal graph model after intervention; Process (b) blocking backdoor pathway: for and Intervention, i.e. and Cross-domain user activity and user domain activity Cross-domain preference representation Intra-domain preference representation The conditional dependency between them is severed, and their conditional probability degenerates into prior probability: Process (c) based on causal mechanism invariance: from post-intervention preference representation and confounding factors to recommendation outcome The causal mechanism remains unchanged; process (d) calculates the causal effect: Measuring preferences indicates the causal effect on recommendation outcomes. and The mathematical expectation representing the confounding factor;

[0091] Ultimately, with This represents the final recommendation result in the X domain, where... Linear fusion is employed; the binary cross-entropy loss function for the X domain is defined as follows:

[0092]

[0093] in, Indicates the predicted recommendation result Chinese users The probability of interacting with an item in domain X. Indicates user Real-world interaction tags with items in the X domain. Let X be the loss function for the domain X. By minimizing this loss function, the probability of predicting user behavior in the domain X is obtained. As close as possible to real interactive labels ;

[0094] The loss function is obtained in the same way in the Y domain. Then the cross-domain sequence recommendation loss function is The accuracy of cross-domain recommendation results can be improved by optimizing the cross-domain sequence recommendation loss function.

[0095] Design sophisticated and effective cross-domain scenario prompts: First, assign the large model the role of a recommendation expert to assist in generating semantic descriptions of missing items; second, input the item name, original attributes, and the top five frequently co-occurring items and their frequencies; finally, guide the large model to infer complementary functions that may attract users when the co-occurrence frequency is higher than five times, and only perform scenario-related descriptions when the co-occurrence frequency is less than five times.

[0096] The temporal-domain biconditional hybrid expert mechanism, unlike intra-domain learning, aims to address the timescale inaccuracy caused by the domain-specific rate of preference change in cross-domain sequences. Its implementation includes: modeling relative time intervals based on the hybrid domain sequences; introducing domain identifiers and absolute timestamps into the gating network to calibrate the timescale; and finally, through collaborative learning and weighting by multiple experts, outputting a unified cross-domain temporal embedding, thereby enhancing the interpretability of the cross-domain interest transfer process.

[0097] Causally reinforced preference learning: To identify the impact of two confounding factors, intra-domain activity and cross-domain activity, on the recommendation model, a structural causal model (SCM) is used to reveal the spurious associations generated through the path of "true preference ← activity → recommendation result". At the same time, to quantify the dynamic changes in user activity, a normalized activity score is calculated with a 90-day time window, and then the original preference is calibrated through a backdoor adjustment formula to eliminate the confounding bias caused by activity.

[0098] To demonstrate the effectiveness of the proposed method (ExcelRec), comparisons were made with existing works on single-domain and cross-domain sequence recommendation. All models were evaluated based on two representative Top-N recommendation metrics: Normalized Discount Cumulative Gain (NDCG) and Hit Rate (HR). These ranking metrics were calculated based on the top 10 predictions for each user in the test set. Experimental results show that the proposed recommendation method achieves state-of-the-art performance on four datasets (Amazon Movies-Books, Clothing-Appliances). Figure 1 Table 1, Table 2 and Figure 2 As shown.

[0099] Table 1. Comparison of recommendation performance between Amazon Movies and Books datasets.

[0100]

[0101] Table 2. Comparison of recommendation performance between Amazon Apparel and Appliances datasets.

[0102]

[0103] In existing work on single-domain sequence recommendation, its poor performance is mainly due to two factors: 1) data sparsity and 2) insufficient semantic modeling. In contrast, various cross-domain sequence recommendation methods generally outperform single-domain recommendation models, verifying that introducing cross-domain data can effectively alleviate the data sparsity problem. Notably, cross-domain recommendation methods based on traditional collaborative filtering lag behind large language model recommendation methods, highlighting the importance of semantic information. Our proposed model further promotes the development of cross-domain recommendation by explicitly modeling the interest evolution process and incorporating auxiliary information enhanced by a large language model. This model effectively addresses the limitations of existing methods (e.g., URLLM lacks causal analysis and temporal learning) through a causal-enhanced preference learning mechanism, achieving state-of-the-art performance in both single-domain and cross-domain sequence recommendation benchmarks.

Claims

1. A cross-domain sequence recommendation method based on causal inference and preference evolution, characterized in that, The steps include the following: Step 1: Obtain user intra-domain interaction sequences and cross-domain interaction sequences, and calculate cross-domain co-occurrence frequency; Step 2: Design prompt word templates based on cross-domain co-occurrence frequency, and enhance the semantic information of items using a large language model to obtain the semantic embedding vector of items; The prompt word template is specifically designed as follows: The large language model is required to combine the item's own information with cross-domain co-occurrence relationships to generate a description; for high-frequency item co-occurrence pairs with a cross-domain co-occurrence frequency exceeding a set threshold, the large language model is instructed to mine their shared usage scenarios and complementary attributes; for low-frequency item co-occurrence pairs with a cross-domain co-occurrence frequency not exceeding a set threshold, the large language model is instructed to analyze their scenario-related descriptions; the large language model integrates all the above information to generate enhanced semantic information for the item, not exceeding the maximum character limit, and generates an item semantic embedding vector based on this enhanced semantic information. ; Step 3: Based on the semantic embedding vectors of items, domain preferences are obtained by learning domain sequence preferences through a domain-specific interest evolution model; Cross-domain preferences are obtained through cross-domain sequence preference learning using a temporal-domain dual-conditional expert mechanism. Step 4: For both intra-domain and cross-domain interaction sequences, use a fixed time window to count the total number of interactions within each time window. Using the global maximum value among all time windows as the standard, normalize the number of intra-domain and cross-domain interactions respectively. The normalized results are used as the user's dynamically changing intra-domain activity and cross-domain activity. Step 5: Perform causal bias removal based on intra-domain preferences, cross-domain preferences, intra-domain activity, and cross-domain activity to generate recommendations; The cross-domain interaction sequence consists of two parts: one is the cross-domain behavior sequence of user interaction with cross-domain items, and the other is the cross-domain time sequence composed of the time of each interaction. Based on the time-domain dual-condition hybrid expert mechanism, the cross-domain behavior sequence and the cross-domain time sequence are extracted and modeled from the cross-domain interaction sequence, and then the dynamic user cross-domain preferences are learned. Specifically, firstly, the semantic embedding vector of the item is... After linear transformation to adapt to the recommendation task, the formula is as follows: in, Represents the semantic embedding vector of an item. The Middle The embedding vector of an item, Indicates the first After the above linear transformation, each item is used as a global embedding representation for the recommendation task. and These are the weight matrices that change linearly at different levels; and These are the bias vectors for different levels of linear transformation; a self-attention mechanism is used to model the cross-domain behavior sequence to obtain the cross-domain behavior embedding. The original cross-domain time embedding is obtained by using time-series coding on the cross-domain time series. ; For each absolute timestamp in the cross-domain time series, an absolute time embedding is learned. A domain embedding is generated using a binary domain identifier, with 0 representing the source domain and 1 representing the target domain. The domain embedding and the absolute time embedding are fused as the gating weights of a time-domain dual-conditional hybrid expert mechanism. Through weighted fusion by experts, a cross-domain time embedding adjusted on a unified time scale is finally output. Cross-domain behavioral embeddings are then incorporated. and adjusted cross-domain time embedding The cross-domain preference representation is obtained by splicing. .

2. The cross-domain sequence recommendation method based on causal inference and preference evolution according to claim 1, characterized in that, The cross-domain co-occurrence frequency is calculated as follows: cross-domain co-occurrence relationships are mined using a large language model to obtain the functional associations between items; when item A and item B appear together in any user sequence, the corresponding cross-domain co-occurrence frequency increases.

3. The cross-domain sequence recommendation method based on causal inference and preference evolution according to claim 1, characterized in that, The item's own information includes the item name and the item's original category; the cross-domain co-occurrence relationship is a list of the top 5 items with the highest cross-domain co-occurrence frequency.

4. The cross-domain sequence recommendation method based on causal inference and preference evolution according to claim 1, characterized in that, The domain interaction sequence consists of two parts: first, the domain behavior sequence of user interaction with items in the domain; and second, the domain time sequence consisting of the time of each interaction. The domain-specific interest evolution model extracts and models domain-specific behavior sequences and domain-specific time sequences from domain-specific interaction sequences, thereby learning dynamic user domain-specific preferences. Specifically, this refers to: semantic embedding vectors of items Principal component analysis is performed for dimensionality reduction to extract low-dimensional item embedding vectors that retain their dominant semantic features. Each item in the user's in-domain behavior sequence is represented by its corresponding low-dimensional item embedding vector. An encoder based on the SASRec architecture is used to model the user's in-domain behavior sequence through a self-attention mechanism, generating in-domain behavior embeddings. The formula for the self-attention mechanism is as follows: Where Q, K, and V represent the query embedding representation, key embedding representation, and value embedding representation, respectively; d represents the embedding dimension. This is a self-attention mechanism; For activation functions; Based on the intra-domain time series in the intra-domain interaction sequence First, calculate the relative time interval in the X domain: By processing relative time intervals through time-series coding, corresponding intra-domain time embeddings are generated. The timing coding formula is as follows: Set the hyperparameter to 1000; This is the time transformation matrix; Embedding behavior within the splicing domain Intradomain time embedding Obtain the in-domain preference representation For another domain Y, its in-domain preference representation It is obtained through exactly the same steps.

5. The cross-domain sequence recommendation method based on causal inference and preference evolution according to claim 1, characterized in that, The adjusted cross-domain time embedding acquisition process is as follows: Each absolute timestamp in a cross-domain time series is decomposed into year, month, and date. A unified time embedding is constructed through a three-layer embedding structure of year embedding, month embedding, and date embedding. in, , , These represent the year embedding vector, month embedding vector, and date embedding vector, respectively. and These are the weight matrices for different levels of linear transformation; and These are the bias vectors for different levels of linear transformation; For computed output, a uniform temporal embedding; Adopting binary domain identifiers As a semantic anchor, Represents the source domain. Represent the target domain, and assign each binary identifier Mapping to Domain Embedding Attention mechanism Embedding the domain in a linear transformation manner and absolute time embedding Fusion; The specific calculation method for gating weights is as follows: in, and These are the weight matrix and the bias vector, respectively. Indicates the gating weight; The cross-domain temporal information representation adjusted by the computational time-domain biconditional hybrid expert mechanism is obtained by weighted fusion of the cross-domain temporal embedding representations of multiple experts, as shown in the following formula: in, It is the first The weight parameters of an expert network, yes Assigned to the The gating weight of each expert, yes The Middle Cross-domain time embedding representation for each user It is the first A cross-domain temporal embedding representation of a user after adjustment by a time-domain biconditional expert mechanism.

6. The cross-domain sequence recommendation method based on causal inference and preference evolution according to claim 1, characterized in that, The aforementioned causal debiasing is a causal-enhanced preference learning method that decomposes user intra-domain preferences and user cross-domain preferences through a backdoor adjustment method, decoupling them from user intra-domain activity and cross-domain activity. In cross-domain sequence recommendation scenarios, there are two confounding factors: user domain activity. and user cross-domain activity ; The confounding factor introduces spurious correlations through two biased paths: one is the activity biased path. , As a representation of in-domain preferences, the activity-biased path improves the recommendation accuracy for highly active users; The second is the deviation path of cross-domain top users. , The cross-domain preference representation causes the recommendation results to be overly biased towards users with frequent interactions; a backdoor adjustment strategy is introduced to correct this bias, and the causal effect of the correction is estimated as follows: In this process (a), based on Bayes' theorem, the intervention conditional probability is transformed into a summation of conditional probabilities based on observation data. Indicates to and Causal graph model after intervention; Process (b) blocking backdoor pathway: for and Intervention, i.e. and Cross-domain user activity and user domain activity Cross-domain preference representation Intra-domain preference representation The conditional dependency between them is severed, and their conditional probability degenerates into prior probability: Process (c) based on causal mechanism invariance: from post-intervention preference representation and confounding factors to recommendation outcome The causal mechanism remains unchanged; process (d) calculates the causal effect: Measuring preferences indicates the causal effect on recommendation outcomes. and The mathematical expectation representing the confounding factor; Ultimately, with This represents the final recommendation result in the X domain, where... Linear fusion is employed; the binary cross-entropy loss function for the X domain is defined as follows: in, Indicates the predicted recommendation result Chinese users The probability of interacting with an item in domain X. Indicates user Real-world interaction tags with items in the X domain. Let X be the loss function for the domain X. By minimizing this loss function, the probability of predicting user behavior in the domain X is obtained. As close as possible to real interactive labels ; The loss function is obtained in the same way in the Y domain. Then the cross-domain sequence recommendation loss function is The accuracy of cross-domain recommendation results can be improved by optimizing the cross-domain sequence recommendation loss function.

Images