User behavior sequence-based interpretable recommendation generation method and recommendation system

Abstract

The invention provides an interpretable recommendation generation method and recommendation system based on a user behavior sequence, and the method comprises the steps: obtaining a historical behavior sequence of a user, the historical behavior sequence comprising a plurality of interaction articles arranged according to a time sequence; on the basis of the historical behavior sequence of the user, personalized natural language explanation for the target article is generated through a sequence perception explanation generator, and the sequence perception explanation generator generates explanation reflecting dynamic evolution of user preferences by deeply modeling the user behavior sequence; and inputting the generated personalized natural language explanation into an explanation enhancement recommender to evaluate the utility of the explanation for recommending the target item. Through a utility-driven evaluation framework constructed by the interpretation enhancement recommender, interpretation quality evaluation is converted from traditional text similarity comparison to objective measurement for improvement of actual recommendation performance, the real value and credibility of interpretation are scientifically verified, and finally, the sense of trust and satisfaction of a user to a recommendation system are effectively enhanced.

Description

Technical Field

[0001] This application relates to the field of recommender system technology, specifically an interpretable recommender generation method and recommender system based on user behavior sequences. Background Technology

[0002] Recommender systems have become a core component of modern online services such as e-commerce, social media, and content platforms, aiming to help users discover items of interest from massive amounts of information. However, many advanced recommendation models, especially those based on deep learning, have complex and opaque internal decision-making processes, resembling a "black box," which severely undermines users' trust in the recommendation results.

[0003] To enhance the transparency and credibility of systems, explainable recommendation techniques have emerged, aiming to provide human-understandable reasons or explanations for recommendation results. In recent years, with the rapid development of large language models, leveraging their powerful natural language generation capabilities to create more fluent and informative recommendation explanations has become an important research direction in this field.

[0004] Existing interpretable recommendation methods based on large language models, whether user / item ID-based, embedding vector-based, or hybrid methods incorporating multiple signals, while improving the quality of interpretation to some extent, still suffer from two fundamental limitations: First, existing methods generally overlook the dynamic sequential nature of user behavior. Most existing methods rely on static user profiles or isolated user-item interaction information to generate explanations. They fail to effectively capture the dynamic trajectory of user preferences evolving over time. For example, a user's interest might gradually shift from "action movies" to "science fiction thrillers." An explanation faithful to their decision-making logic should reflect this dynamic "leap" in preference. However, explanations generated by existing technologies often degenerate into mere restates of the static attributes of the target item, resembling a "post-hoc justification" rather than a dynamic presentation of the user's true decision-making logic. This significantly limits the personalization level of the explanation and the user's ultimate sense of identification.

[0005] Second, the evaluation paradigm is severely disconnected from the true utility of the explanation. Existing technologies generally use text similarity-based metrics such as BLEU and BERTScore to evaluate the quality of explanations. These metrics only focus on the overlap between the generated explanation and the reference text (usually user reviews) in terms of surface vocabulary or semantics, but fail to measure whether the explanation truly helps users make better decisions. More seriously, using user reviews as the "gold standard" for explanations is itself a "paradigm misalignment." User reviews often describe the universal characteristics of an item, while a good recommendation explanation should focus on the personalized decision-making motivations of users when making choices.

[0006] Therefore, how to overcome the above-mentioned shortcomings of existing technologies and propose an explainable recommendation method that can capture users' dynamic behavior, generate high-fidelity personalized explanations, and scientifically evaluate their true utility is a problem that urgently needs to be solved in the current technology field. Summary of the Invention

[0007] The problem this invention addresses is how to capture user dynamic behavior and generate high-fidelity personalized interpretations.

[0008] To address the aforementioned problems, this invention provides an interpretable recommendation generation method based on user behavior sequences, an interpretable recommendation system based on user behavior sequences, an electronic device, and a storage medium.

[0009] In a first aspect, the present invention provides an interpretable recommendation generation method based on user behavior sequences, comprising: Obtain the user's historical behavior sequence, which includes multiple interactive items arranged in chronological order; Based on the user's historical behavior sequence, a personalized natural language explanation for the target item is generated through a sequence-aware explanation generator, wherein the sequence-aware explanation generator generates an explanation that reflects the dynamic evolution of user preferences by deeply modeling the user behavior sequence. The generated personalized natural language explanation is input into the explanation-enhanced recommender to evaluate the utility of the explanation in recommending the target item, wherein the explanation-enhanced recommender evaluates the utility of the explanation by quantifying the change in recommendation performance before and after the explanation is introduced.

[0010] Optionally, based on the user's historical behavior sequence, a personalized natural language explanation for the target item is generated through a sequence-aware explanation generator, specifically including: The historical behavior sequence is encoded using a unified dual-path sequence encoding method that combines behavior and semantics to obtain a behavior sequence embedding representation and a semantic sequence embedding representation. Based on the expert hybrid mechanism, dynamic information fusion and semantic space adaptation are performed on the behavior sequence embedding representation and the semantic sequence embedding representation to obtain the adapted behavior embedding vector and semantic embedding vector. The adapted behavior embedding vector and semantic embedding vector are injected into the large language model so that the large language model generates the personalized natural language interpretation.

[0011] Optionally, the step of performing unified behavior and semantic dual-path sequence encoding on the historical behavior sequence to obtain behavior sequence embedding representation and semantic sequence embedding representation includes: The item ID sequence in the historical behavior sequence is encoded using a pre-trained sequence recommendation model to obtain the behavior sequence embedding representation; A sentence encoder is used to encode the text description sequence corresponding to each item in the historical behavior sequence to obtain a semantic sequence embedding representation.

[0012] Optionally, the adapted behavior embedding vector and semantic embedding vector are injected into the large language model, including: The adapted behavior embedding vector and semantic embedding vector are concatenated with the system prompt embedding and the category text embedding sampled from the item sequence to form the input sequence of the large language model; The adapted behavior embedding vector and semantic embedding vector are directly added to the query, key, and value matrix of each attention module of the large language model.

[0013] Optionally, the generated personalized natural language explanation is input into an explanation-enhanced recommender to evaluate the utility of the explanation in recommending the target item, including: The personalized natural language interpretation is encoded into an interpretation vector using a semantic encoder; The explanatory vector is input into the hypernetwork to generate a set of adapter parameters for adjusting the parameters of the downstream recommendation model; Using a downstream recommendation model with adjusted parameters, the user's preference score for the target item is predicted based on the user's historical behavior sequence and the explanatory vector. The recommendation utility of the personalized natural language interpretation is quantified based on the changes in the recommendation model's performance metrics before and after the introduction of the interpretation.

[0014] Optionally, the downstream recommendation model is a Transformer-based sequence recommendation model; the adapter parameters are injected into the corresponding layer of the downstream recommendation model through an addition operation.

[0015] Optionally, the explanation-enhanced recommender is optimized during the training phase using the Bayesian Personalized Ranking (BPR) loss function, where model predictions are conditioned on the user's historical behavior sequence and the embedding vector of the gold-standard explanation.

[0016] Secondly, the present invention provides an interpretable recommendation system based on user behavior sequences, comprising: A sequence-aware explanation generator is used to acquire a user's historical behavior sequence and generate a personalized natural language explanation for a target item based on the historical behavior sequence. The sequence-aware explanation generator is configured to generate an explanation that reflects the dynamic evolution of user preferences by deeply modeling the user behavior sequence. An explanation-enhanced recommender, communicatively connected to the sequence-aware explanation generator, is configured to receive the personalized natural language explanation and evaluate the utility of the explanation for recommending the target item, wherein the explanation-enhanced recommender is configured to evaluate the utility of the explanation by quantifying the change in recommendation performance before and after the explanation is introduced.

[0017] Optionally, the dual-path sequence encoding module is used to perform behavior sequence encoding and semantic sequence encoding on the user's historical behavior sequence to obtain behavior sequence embedding representation and semantic sequence embedding representation, respectively; An adapter module based on expert hybridization is used to fuse and adapt the behavior sequence embedding representation and the semantic sequence embedding representation; The explanation generation module includes a large language model, which is configured to receive adapted embedding vectors and generate the personalized natural language explanation.

[0018] Thirdly, the present invention provides an electronic device including a processor, a communication interface, a memory, and a bus, wherein the processor, the communication interface, and the memory communicate with each other through the bus, and the processor can call logical instructions in the memory to execute the steps of the method provided in the first aspect.

[0019] Fourthly, the present invention provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the interpretable recommendation generation method based on user behavior sequences as described in the first aspect.

[0020] The beneficial effects of the interpretable recommendation generation method based on user behavior sequences of the present invention are as follows: By using a sequence-aware explanation generator to deeply model the dynamic evolution of user behavior sequences, high-quality natural language explanations that faithfully reflect users' personalized decision-making logic can be generated, significantly improving the transparency of recommendation decisions. At the same time, by constructing a utility-driven evaluation framework for the explanation-enhanced recommender, the evaluation of explanation quality is transformed from traditional text similarity comparison to an objective measure of its actual recommendation performance improvement, scientifically verifying the true value and credibility of the explanation, and ultimately effectively enhancing users' trust and satisfaction with the recommendation system. Attached Figure Description

[0021] Figure 1 This is a flowchart of an interpretable recommendation generation method based on user behavior sequences in an embodiment of the present invention; Figure 2 This is a schematic diagram of the sequence-aware interpretation generator structure in an embodiment of the present invention; Figure 3 This is a schematic diagram illustrating the structure of the enhanced recommender in an embodiment of the present invention; Figure 4This is a diagram showing the ablation experiment results of the sequence sensing component in the SELLER framework in this embodiment of the invention; Figure 5 This is a structural block diagram of the electronic device in an embodiment of the present invention. Detailed Implementation

[0022] To better understand the purpose, technical solution, and advantages of this application, the application is described and explained below in conjunction with the accompanying drawings and embodiments.

[0023] Unless otherwise defined, the technical or scientific terms used in this application shall have the general meaning understood by one of ordinary skill in the art to which this application pertains. Words such as “a,” “an,” “an,” “the,” “the,” and “these” used in this application do not indicate quantitative limitation and may be singular or plural. The terms “comprising,” “including,” “having,” and any variations thereof used in this application are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or device that comprises a series of steps or modules (units) is not limited to the listed steps or modules (units) but may include steps or modules (units) not listed, or may include other steps or modules (units) inherent to these processes, methods, products, or devices. Words such as “connected,” “linked,” and “coupled” used in this application are not limited to physical or mechanical connections but may include electrical connections, whether direct or indirect. “Multiple” used in this application refers to two or more. “And / or” describes the relationship between related objects, indicating that three relationships may exist; for example, “A and / or B” can represent: A alone, A and B simultaneously, and B alone. Normally, the character " / " indicates that the objects before and after it are in an "or" relationship. The terms "first," "second," "third," etc., used in this application are merely to distinguish similar objects and do not represent a specific order of objects.

[0024] This invention aims to address the problems of existing technologies neglecting the dynamic nature of user behavior and the disconnect between evaluation paradigms and actual utility. Specifically, this invention aims to achieve the following objectives: This paper proposes an explanation generation method capable of deeply modeling the dynamic evolution of user preferences. By uniformly capturing behavioral patterns and semantic evolution in user interaction sequences, this method generates highly personalized natural language explanations that faithfully reflect the user's decision-making logic, thereby significantly improving the transparency of the recommendation decision-making process.

[0025] This paper proposes a unified, utility-driven evaluation framework. This framework aims to transform the evaluation of explanation quality from a traditional "text matching" problem into an objective "utility measurement" problem. Specifically, it scientifically measures the true value of explanations by quantifying their actual impact on the performance of downstream recommendation tasks, thereby effectively verifying the credibility of recommendation decisions.

[0026] Ultimately, by improving the transparency and credibility of decision-making, a stronger, more explainable recommendation system can be built, thereby enhancing user trust and satisfaction.

[0027] The overall framework of this invention (hereinafter referred to as the SELLER framework) consists of two core components: A sequence-aware interpretation generator is responsible for deep modeling of user behavior sequences to address the challenge of decision transparency. Its core function is to generate personalized natural language interpretations that faithfully reflect the dynamic evolution of user preferences.

[0028] This paper presents an explanation-enhanced recommender, which is a unified evaluation framework designed to address the challenge of decision credibility. It scientifically evaluates the true utility and value of explanations by quantifying their actual impact on the performance of downstream recommendation tasks.

[0029] First, define the core entities and symbols involved in this invention, and let... Represents a set of users. Represents a collection of items. For any user Its interaction history is defined as a sequence of user behaviors arranged in chronological order. , denoted as: ; in, It is the user at the time step Interacted items. This sequence is a dynamic trajectory of user preferences, fundamental to understanding their decision-making logic. Furthermore, each item... Each is associated with a text description (For example, product attributes, movie synopses, etc.) provide explicit semantic information for modeling user preferences.

[0030] Based on the above definition, the technical solution of the present invention includes the following two core tasks: Task 1: Sequence-aware interpretation generation (addressing the transparency challenge); The primary objective of this invention is to generate natural language explanations that faithfully reflect user behavior sequences for recommendation decisions. Formally, for a given user-item pair... The goal is to learn a generating function. This function can be based on the user's complete sequence of actions. And related semantic information, to generate a natural language explanation. : ; The core of this definition is that it mandates that the explanation generation process must be deeply dependent on the user's behavioral sequence, ensuring that its output can truly reveal the dynamic logic behind the recommendation decision.

[0031] Task 2: Explain the utility assessment (addressing credibility challenges); To verify the credibility of the explanations, this invention further defines an explanation utility evaluation framework. The core idea of ​​this framework is that a credible explanation must have positive utility value for the recommendation task it explains. To this end, this invention introduces an Explanation Enhancement Recommender (EER), denoted as the function... When predicting a user's preference score for an item, it will simultaneously use the user's behavior sequence. and the explanation generated As a condition: ; By comparing the changes in recommendation performance before and after the introduction of explanations, the recommendation utility of explanations can be quantitatively evaluated, thus transforming the "credibility" of explanations from a subjective text matching problem into an objective utility measurement problem.

[0032] like Figure 1 As shown in the embodiment of the present invention, an interpretable recommendation generation method based on user behavior sequences includes: S1. Obtain the user's historical behavior sequence, which includes multiple interactive items arranged in chronological order; S2. Based on the user's historical behavior sequence, a personalized natural language explanation for the target item is generated through a sequence-aware explanation generator, wherein the sequence-aware explanation generator generates an explanation that reflects the dynamic evolution of user preferences by deeply modeling the user behavior sequence. Based on the user's historical behavior sequence, a personalized natural language explanation for the target item is generated through a sequence-aware explanation generator, specifically including: The historical behavior sequence is encoded using a unified dual-path sequence encoding method that combines behavior and semantics to obtain a behavior sequence embedding representation and a semantic sequence embedding representation. Based on the expert hybrid mechanism, dynamic information fusion and semantic space adaptation are performed on the behavior sequence embedding representation and the semantic sequence embedding representation to obtain the adapted behavior embedding vector and semantic embedding vector. The adapted behavior embedding vector and semantic embedding vector are injected into the large language model so that the large language model generates the personalized natural language interpretation.

[0033] The step of performing unified dual-path sequence encoding on the historical behavior sequence to obtain behavior sequence embedding representation and semantic sequence embedding representation includes: The item ID sequence in the historical behavior sequence is encoded using a pre-trained sequence recommendation model to obtain the behavior sequence embedding representation; A sentence encoder is used to encode the text description sequence corresponding to each item in the historical behavior sequence to obtain a semantic sequence embedding representation.

[0034] Injecting the adapted behavior embedding vector and semantic embedding vector into the large language model includes: The adapted behavior embedding vector and semantic embedding vector are concatenated with the system prompt embedding and the category text embedding sampled from the item sequence to form the input sequence of the large language model; The adapted behavior embedding vector and semantic embedding vector are directly added to the query, key, and value matrix of each attention module of the large language model.

[0035] Specifically, such as Figure 2 As shown, the goal of the sequence-aware explanation generator is to generate high-quality, personalized explanations that can accurately capture the evolution of user preferences.

[0036] A user's decision-making logic is rooted in their complete interaction history, and a single-dimensional information source is insufficient. The user's behavioral sequence itself contains rich collaborative filtering signals, while the textual description of items reveals the explicit semantic themes of the user's interests. The two are highly complementary. Based on this, this invention designs a dual-path sequence encoding mechanism that unifies behavior and semantics.

[0037] Behavior sequence encoding: A pre-trained sequence recommendation model (e.g., SASRec) is used as the behavior sequence encoder. Given a user's sequence of actions The encoder converts it into an embedding matrix. This matrix captures temporal patterns and collaborative information in user interaction behavior: ; in, It is the dimension of the embedded vector.

[0038] Semantic sequence encoding: Utilizing a sentence encoder (e.g., Sentence-T5) as the semantic sequence encoder. The encoder is responsible for processing sequences of user actions. The text description sequence corresponding to each item And generate a semantic embedding matrix. : ; This matrix explicitly models the evolution trajectory of user interests in the semantic space.

[0039] To effectively integrate two heterogeneous sequence representations, behavioral and semantic, and align them with the semantic space of a large language model (LLM), this invention introduces a parallel sequence adapter based on expert hybridization (MoE).

[0040] For each path The sequence embedding matrix is ​​first processed by a Transformer encoder to integrate temporal information, and then average pooling is used to obtain the overall sequence-level representation. Subsequently, this representation is fed into the MoE adapter for final transformation and alignment to obtain the adapted embedding vector. : ; in, It is the number of experts. Is the decision of the first A gating network with expert weights, each... It is a feedforward network. This process ultimately produces two depth-fitted embedding vectors: (representing behavioral patterns) and (Represents semantic features).

[0041] To ensure that the large language model can fully utilize the extracted sequence information when generating explanations, this invention not only injects the adapted sequence embedding vector into the input layer of the model, but also deeply integrates it into the model's attention calculation process, so as to faithfully generate explanations based on user behavior sequences.

[0042] Specifically, the input text sequence of the model is constructed as follows: ; in, It's an embedded system prompt. and It is the adapted sequence embedding. It is a category text embedding sampled from the item sequence.

[0043] Furthermore, the two adapted embedding vectors and Add directly to the query, key, and value matrices of each attention module in the LLM: ; in, It is the LLM number The original matrix of the layer. This design allows the model to dynamically adjust its attention distribution based on the user's historical behavior patterns.

[0044] During the training phase, the backbone parameters of the LLM are frozen, and only the newly introduced adapter module is fine-tuned. SEG is optimized using a standard language modeling objective, with its loss function... The definition is as follows: ; in, It is the first in the target interpretation sequence Each token.

[0045] S3. Input the generated personalized natural language explanation into the explanation enhancement recommender to evaluate the utility of the explanation in recommending the target item, wherein the explanation enhancement recommender evaluates the utility of the explanation by quantifying the change in recommendation performance before and after the explanation is introduced.

[0046] The generated personalized natural language explanation is input into the explanation-enhanced recommender to evaluate the utility of the explanation in recommending the target item, including: The personalized natural language interpretation is encoded into an interpretation vector using a semantic encoder; The explanatory vector is input into the hypernetwork to generate a set of adapter parameters for adjusting the parameters of the downstream recommendation model; Using a downstream recommendation model with adjusted parameters, the user's preference score for the target item is predicted based on the user's historical behavior sequence and the explanatory vector. The recommendation utility of the personalized natural language interpretation is quantified based on the changes in the recommendation model's performance metrics before and after the introduction of the interpretation.

[0047] In this embodiment, the downstream recommendation model is a Transformer-based sequence recommendation model; the adapter parameters are injected into the corresponding layer of the downstream recommendation model through an addition operation.

[0048] In this embodiment, the explanation-enhanced recommender is optimized using the Bayesian Personalized Ranking (BPR) loss function during the training phase, where the model prediction is conditioned on the user's historical behavior sequence and the embedding vector of the gold standard explanation.

[0049] like Figure 3 As shown, the core of the explanation-enhanced recommender is to scientifically evaluate the true utility of the explanation by quantifying its actual impact on downstream recommendation tasks.

[0050] Given a user-item pair by SEG The generated explanation First, a semantic encoder is used. Encode it into a dense interpretation vector : ; The explanatory vector is then fed into a hypernetwork. This hypernetwork will then process each Transformer layer of the downstream recommendation model (e.g., SASRec). Generate a specific set of adapter parameters : ; These dynamically generated adapter parameters will be injected into the corresponding layer of SASRec (e.g., via addition operations). This allows the recommendation behavior to be finely controlled by the explanatory content, thus realizing the transformation of the recommendation model from static to dynamic.

[0051] During the independent training phase, the augmented recommender is optimized using a standard Bayesian Personalized Ranking (BPR) loss function. Its core principle is that the model, when making ranking predictions, must simultaneously consider the user's historical behavior sequence. Embedded vectors as interpreted by the "gold standard" As a condition: ; in, These are positive samples (items that the user has actually interacted with). It is a negative sample. It is the predicted score of EER. It is the sigmoid function.

[0052] During the evaluation phase, the embedding vectors corresponding to the explanations generated by different explanation models will be used. As input, we examine the changes in EER recommendation performance. If an explanation can improve the accuracy of recommendations (e.g., Recall@K, NDCG@K), it proves that the explanation has successfully captured the user's real, personalized decision-making motivations, and therefore has higher recommendation utility and credibility.

[0053] To better understand this solution, the following example is provided to illustrate the entire process by which this invention generates high-quality explanations from user history and guides recommendations: Input user's historical behavior sequence: The system obtains a user's historical interaction item list, which includes a series of bars and restaurants, such as 'Hotel Tango', 'Union 50', 'HopCat - St.Louis', 'The Inferno Room', 'Workingman's Friend', etc.

[0054] The Sequence-Aware Interpretation Generator (SEG) processes the data. Data preprocessing: The data processor inputs the historical interaction list into the SEG.

[0055] Dual-path encoding: The SEG processes the sequence in parallel. The behavioral sequence encoder captures user preference patterns for places such as "bars" and "restaurants" from the ID sequence. The semantic sequence encoder processes the textual descriptions of these places and combines them with "sampled category texts" such as ['Distilleries', 'Nightlife', 'Wine & Spirits'], ['Arts & Entertainment', 'Music Venues', 'American (New)'] to extract the deeper semantics of user interests.

[0056] Generated Explanations: Based on the deeply fused sequence information, the Large Language Model (LLM) generated the following three key explanations reflecting the evolution of user preferences (LLM Output Explanation): "Bars, bars catering to a wide range of tastes" "Restaurants, family-friendly dining options in my area" "American (Traditional) Places that offer a taste of authentic, down-to-earth American fare." Explanation of Enhanced Recommender (EER) for utility verification and recommendation: Generating a recommendation list: EER takes the three interpretations generated by SEG as input. These interpretations successfully capture the key categories of the target item "Downtown Olly's" (whose actual category list includes 'American (Traditional)', 'Restaurants', and 'Bars').

[0057] Utility verification: Based on these high-quality explanations, EER successfully included the target item "Downtown Olly's" in its generated Top 5 recommendation list.

[0058] Performance Comparison: In contrast, a standard recommendation model without the explanation enhancements of this invention (w / o HGA RecItem List, such as SASRec) generates a recommendation list that contains no target items at all, and its recommended content (such as 'HillStreet Grill', 'Smashburger', etc.) deviates significantly from the user's dynamic interests.

[0059] Extensive experiments on two large-scale public datasets (Yelp and KuaiRec) demonstrate that, compared to existing technologies, it has the following significant advantages and positive effects. Specific experimental data and analysis are as follows: It demonstrates superior recommendation utility and significantly enhances decision credibility. To verify the true utility (credibility) of the explanations, this invention evaluates the downstream recommendation performance of explanations generated by different methods within a unified Explanation Enhancement Recommender (EER) framework. As shown in Table 1, the experimental results clearly demonstrate: This invention (SELLER) demonstrates superior recommendation utility across the board: Explanations generated using this invention achieve the best downstream recommendation performance across all evaluation metrics on both datasets. Compared to the top-performing baseline model XRec, this invention achieves a significant relative improvement of 27.52% on the NDCG@10 metric of the Yelp dataset and a relative improvement of 6.57% on the KuaiRec dataset. This strongly demonstrates that, through its deep modeling capabilities of user behavior sequences, this invention can generate explanations that more accurately reflect the true motivations behind user decisions, thus providing the most informative and valuable semantic signals for recommendation tasks.

[0060] The necessity of sequence modeling was verified: a key phenomenon is that on the KuaiRec dataset, where interaction sequences are denser, methods that failed to effectively capture sequence dynamics (PETER and PEPLER) performed even worse in recommendation utility than the Empty baseline, which provided no explanation. This indicates that in complex scenarios, poor-quality explanations that fail to faithfully reflect dynamic user preferences are not only unhelpful but may also introduce noise, negatively interfering with recommendation decisions. This result, conversely, confirms the extreme importance of the user behavior sequence-centric design of this invention for generating reliable explanations.

[0061] Table 1: Results of the evaluation of the recommended utility. It is most faithful in terms of generation quality, effectively ensuring the transparency of decision-making.

[0062] In terms of traditional text generation quality dimensions, this invention also demonstrates optimal performance, ensuring the transparency of the decision-making process. As shown in Table 2: The explanations presented in this invention are the most faithful to both lexical and semantic levels: on both datasets, this invention significantly outperforms all baseline models in both BLEU (lexical overlap) and BERTScore (semantic similarity). Its advantages are particularly pronounced on the KuaiRec dataset, which features longer sequences and more complex patterns; for example, the BLEU score is improved by 34.90% relative to the strongest baseline. This fully demonstrates that the dual-path sequence encoding mechanism of this invention can effectively extract key information from the user's interaction history, thereby generating explanatory text that is closer to the "gold standard."

[0063] This study reveals and resolves the "misalignment" problem in the evaluation paradigm: A joint analysis of Tables 1 and 2 reveals a profound phenomenon: although this invention significantly improves text similarity (fidelity) on KuaiRec compared to Yelp, its improvement in recommendation utility (credibility) on Yelp is even more significant. This reveals the misalignment in the existing evaluation paradigm that "higher text similarity does not necessarily equate to higher actual utility." The fundamental reason is that the "gold standard" (such as user reviews) itself may contain a large amount of generic description or noise. Therefore, the evaluation framework advocated by this invention, which focuses on recommendation utility, is more scientific and reliable.

[0064] The method, framework, and system proposed in this invention have been verified through detailed experiments on two large-scale public datasets (Yelp and KuaiRec). Compared with existing technologies, they have the following significant advantages and positive effects, and the specific experimental data and analysis are as follows: It demonstrates excellent recommendation effectiveness and significantly enhances decision credibility. To verify the true utility (credibility) of the explanations, this invention evaluated the downstream recommendation performance of explanations generated by different methods within a unified Explanation Enhancement Recommender (EER) framework. As shown in Table 1, the experimental results clearly demonstrate: This invention (SELLER) demonstrates superior recommendation utility across the board: The explanations generated using this invention achieve the best downstream recommendation performance across all evaluation metrics on both datasets. Compared to the top-performing baseline model XRec, this invention achieves a significant relative improvement of 27.52% on the NDCG@10 metric of the Yelp dataset and a relative improvement of 6.57% on the KuaiRec dataset. This strongly demonstrates that, through its deep modeling capabilities of user behavior sequences, this invention can generate explanations that more accurately reflect the true motivations behind user decisions, thus providing the most informative and valuable semantic signals for recommendation tasks.

[0065] The necessity of sequence modeling was verified: A key phenomenon is that on the KuaiRec dataset, where interaction sequences are denser, methods that failed to effectively capture sequence dynamics (PETER and PEPLER) performed even worse in recommendation utility than the Empty baseline, which provided no explanation. This indicates that in complex scenarios, low-quality explanations that fail to faithfully reflect dynamic user preferences are not only useless but may also introduce noise, negatively interfering with recommendation decisions. This result, conversely, confirms the extreme importance of the user behavior sequence-centric design of this invention for generating reliable explanations.

[0066] Table 2: Explanation of Generative Quality Assessment Results

[0067] Through a series of detailed ablation experiments, such as Figure 4 As shown, the contribution of the core design of this invention was quantitatively evaluated: Dual-path sequence encoding is the cornerstone of success: the experimental variant "w / o DE" (simultaneous removal of behavioral and semantic sequence encoding) experienced the most dramatic performance degradation, which strongly demonstrates the core design motivation of this invention—that is, in order to fully understand user preferences and generate high-quality interpretations, it is necessary to capture both implicit behavioral patterns and explicit semantic features of users at the same time.

[0068] The contributions of different sequence information are scenario-dependent: on the Yelp dataset, the performance degradation caused by removing semantic sequence encoding (w / o SE) is more significant than that caused by removing behavioral sequence encoding (w / o BE), indicating that semantic content has a greater impact in scenarios like Yelp, which are rich in high-quality text reviews. However, in the short video recommendation scenario KuaiRec, the contributions of the two encodings are roughly equal. This finding validates the flexibility and adaptability of the dual-path design of this invention to different scenarios.

[0069] This invention also provides an interpretable recommendation system based on user behavior sequences, the recommendation system comprising: A sequence-aware explanation generator is used to acquire a user's historical behavior sequence and generate a personalized natural language explanation for a target item based on the historical behavior sequence. The sequence-aware explanation generator is configured to generate an explanation that reflects the dynamic evolution of user preferences by deeply modeling the user behavior sequence. An explanation-enhanced recommender, communicatively connected to the sequence-aware explanation generator, is configured to receive the personalized natural language explanation and evaluate the utility of the explanation for recommending the target item, wherein the explanation-enhanced recommender is configured to evaluate the utility of the explanation by quantifying the change in recommendation performance before and after the explanation is introduced.

[0070] The Sequence-Aware Interpretation Generator (SEG) includes: The dual-path sequence encoding module is used to encode the user's historical behavior sequence into behavior sequence encoding and semantic sequence encoding, respectively obtaining behavior sequence embedding representation and semantic sequence embedding representation; An adapter module based on expert hybridization is used to fuse and adapt the behavior sequence embedding representation and the semantic sequence embedding representation; The explanation generation module includes a large language model, which is configured to receive adapted embedding vectors and generate the personalized natural language explanation.

[0071] like Figure 5 As shown in the figure, an electronic device provided by an embodiment of the present invention includes: a processor 610, a communication interface 620, a memory 630, and a communication bus 640, wherein the processor 610, the communication interface 620, and the memory 630 communicate with each other through the communication bus 640. The processor 610 can call logical instructions in the memory 630 to execute the following method: Obtain the user's historical behavior sequence, which includes multiple interactive items arranged in chronological order; Based on the user's historical behavior sequence, a sequence-aware interpretation generator generates a personalized natural language interpretation for the target item. The sequence-aware interpretation generator generates an interpretation that reflects the dynamic evolution of user preferences by deeply modeling the user behavior sequence. The generated personalized natural language explanation is input into the explanation-enhanced recommender to evaluate the utility of the explanation in recommending the target item, wherein the explanation-enhanced recommender evaluates the utility of the explanation by quantifying the change in recommendation performance before and after the explanation is introduced.

[0072] Furthermore, the logical instructions in the aforementioned memory 630 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0073] This invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, is implemented to perform the methods provided in the above embodiments.

[0074] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of various embodiments or some parts of embodiments.

[0075] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An interpretable recommendation generation method based on user behavior sequences, characterized in that, The method includes: Obtain the user's historical behavior sequence, which includes multiple interactive items arranged in chronological order; Based on the user's historical behavior sequence, a personalized natural language explanation for the target item is generated through a sequence-aware explanation generator, wherein the sequence-aware explanation generator generates an explanation that reflects the dynamic evolution of user preferences by deeply modeling the user behavior sequence. The generated personalized natural language explanation is input into the explanation-enhanced recommender to evaluate the utility of the explanation in recommending the target item, wherein the explanation-enhanced recommender evaluates the utility of the explanation by quantifying the change in recommendation performance before and after the explanation is introduced.

2. The method according to claim 1, characterized in that, Based on the user's historical behavior sequence, a personalized natural language explanation for the target item is generated through a sequence-aware explanation generator, specifically including: The historical behavior sequence is encoded using a unified dual-path sequence encoding method that combines behavior and semantics to obtain a behavior sequence embedding representation and a semantic sequence embedding representation. Based on the expert hybrid mechanism, dynamic information fusion and semantic space adaptation are performed on the behavior sequence embedding representation and the semantic sequence embedding representation to obtain the adapted behavior embedding vector and semantic embedding vector. The adapted behavior embedding vector and semantic embedding vector are injected into the large language model so that the large language model generates the personalized natural language interpretation.

3. The method according to claim 2, characterized in that, The step of performing unified dual-path sequence encoding on the historical behavior sequence to obtain behavior sequence embedding representation and semantic sequence embedding representation includes: The item ID sequence in the historical behavior sequence is encoded using a pre-trained sequence recommendation model to obtain the behavior sequence embedding representation; A sentence encoder is used to encode the text description sequence corresponding to each item in the historical behavior sequence to obtain a semantic sequence embedding representation.

4. The method according to claim 2, characterized in that, Injecting the adapted behavior embedding vector and semantic embedding vector into the large language model includes: The adapted behavior embedding vector and semantic embedding vector are concatenated with the system prompt embedding and the category text embedding sampled from the item sequence to form the input sequence of the large language model; The adapted behavior embedding vector and semantic embedding vector are directly added to the query, key, and value matrix of each attention module of the large language model.

5. The method according to claim 1, characterized in that, The generated personalized natural language explanation is input into the explanation-enhanced recommender to evaluate the utility of the explanation in recommending the target item, including: The personalized natural language interpretation is encoded into an interpretation vector using a semantic encoder; The explanatory vector is input into the hypernetwork to generate a set of adapter parameters for adjusting the parameters of the downstream recommendation model; Using a downstream recommendation model with adjusted parameters, the user's preference score for the target item is predicted based on the user's historical behavior sequence and the explanatory vector. The recommendation utility of the personalized natural language interpretation is quantified based on the changes in the recommendation model's performance metrics before and after the introduction of the interpretation.

6. The method according to claim 5, characterized in that, The downstream recommendation model is a Transformer-based sequence recommendation model; the adapter parameters are injected into the corresponding layer of the downstream recommendation model through an addition operation.

7. The method according to claim 1, characterized in that, The explanation-enhanced recommender is optimized during the training phase using the Bayesian Personalized Ranking (BPR) loss function, where model predictions are conditioned on the user's historical behavior sequence and the embedding vector of the gold standard explanation.

8. An interpretable recommendation system based on user behavior sequences, characterized in that, include: A sequence-aware explanation generator is used to acquire a user's historical behavior sequence and generate a personalized natural language explanation for a target item based on the historical behavior sequence. The sequence-aware explanation generator is configured to generate an explanation that reflects the dynamic evolution of user preferences by deeply modeling the user behavior sequence. An explanation-enhanced recommender, communicatively connected to the sequence-aware explanation generator, is configured to receive the personalized natural language explanation and evaluate the utility of the explanation for recommending the target item, wherein the explanation-enhanced recommender is configured to evaluate the utility of the explanation by quantifying the change in recommendation performance before and after the explanation is introduced.

9. The system according to claim 8, characterized in that, The Sequence-Aware Interpretation Generator (SEG) includes: The dual-path sequence encoding module is used to encode the user's historical behavior sequence into behavior sequence encoding and semantic sequence encoding, respectively obtaining behavior sequence embedding representation and semantic sequence embedding representation; An adapter module based on expert hybridization is used to fuse and adapt the behavior sequence embedding representation and the semantic sequence embedding representation; The explanation generation module includes a large language model, which is configured to receive adapted embedding vectors and generate the personalized natural language explanation.

10. A non-transitory computer-readable storage medium, characterized in that, The storage medium stores a computer program that, when executed by a processor, implements the interpretable recommendation generation method based on user behavior sequences as described in any one of claims 1 to 7.

Images