A user interest deep mining and multi-level explainability recommendation method and device

Abstract

This invention discloses a method and apparatus for deep mining of user interests and multi-level interpretable recommendation, comprising the following steps: Step S1, obtaining user interest features based on the user's historical interaction data; Step S2, obtaining a user interest graph based on the user interest features; Step S3, obtaining candidate items that match the user's interests based on the user interest graph and prompt templates; Step S4, obtaining attractive and interpretable recommendation text based on the candidate items. Using the technical solution of this invention, multi-level user interests can be captured and represented, while simultaneously generating interpretable recommendation text.

Description

Technical Field

[0001] This invention belongs to the field of natural language processing technology, and in particular relates to a method and apparatus for deep mining of user interests and multi-level interpretable recommendation. Background Technology

[0002] With the rapid development of the internet, users are facing an increasingly serious problem of information overload. Recommendation systems, as an effective means to solve this problem, are widely used in e-commerce, content distribution, and other fields. However, traditional recommendation systems often only focus on users' explicit interests, making it difficult to delve into their implicit interests, and the recommendation results lack interpretability, leading to low user satisfaction. Summary of the Invention

[0003] The technical problem to be solved by the present invention is to provide a method and apparatus for deep mining of user interests and multi-level interpretable recommendation. By deeply mining users' historical interaction data, the method captures and represents users' multi-level interests by utilizing the text generation capabilities of large models, and at the same time generates interpretable recommendation copy.

[0004] To achieve the above objectives, the present invention adopts the following technical solution:

[0005] A method for deep mining of user interests and multi-level interpretability recommendation includes the following steps:

[0006] Step S1: Obtain user interest characteristics based on the user's historical interaction data;

[0007] Step S2: Obtain the user interest map based on user interest characteristics;

[0008] Step S3: Based on the user interest graph and the prompt template, obtain candidate items that match the user's interests;

[0009] Step S4: Based on the candidate items, obtain attractive and explainable recommendation copy.

[0010] Preferably, in step S1, the user's historical interaction data includes: browsing history, click behavior, and purchase history, and the user's interest characteristics include: the user's explicit interest characteristics and implicit interest characteristics.

[0011] Preferably, step S2 includes:

[0012] Based on user interest features, a multi-level user interest representation model is constructed; the user interest representation model includes multiple levels of interest nodes, each node representing a specific interest point or need.

[0013] By connecting the nodes of interest, a complete user interest graph can be formed.

[0014] Preferably, step S3 includes:

[0015] Based on the user's interest graph and the characteristics of the recommended items, a prompt template is obtained; the prompt template includes the title, description, selling points of the recommended items, as well as keywords and phrases related to the user's interests;

[0016] Based on the prompt template and the large language model, quickly search and match within the existing projects to find candidate items that match the user's interests.

[0017] As an alternative, the method also includes: step S5, continuously optimizing the recommendation process and prompt templates based on user feedback data and behavioral data.

[0018] This invention also provides a device for in-depth user interest mining and multi-level interpretable recommendation, comprising:

[0019] The data mining module is used to obtain user interest characteristics based on the user's historical interaction data;

[0020] The module is used to generate a user interest graph based on user interest characteristics;

[0021] The retrieval module is used to obtain candidate items that match the user's interests based on the user's interest graph and prompt templates;

[0022] The generation module generates attractive and explainable recommendation text based on candidate items.

[0023] As a preferred approach, the user's historical interaction data includes: browsing history, click behavior, and purchase history; and the user's interest characteristics include: the user's explicit interests and implicit interests.

[0024] Preferably, the building modules include:

[0025] The first building unit is used to construct a multi-level user interest representation model based on user interest features; wherein, the user interest representation model includes multiple levels of interest nodes;

[0026] The second building block is used to connect interest nodes to form a user interest graph.

[0027] Preferably, the retrieval module includes:

[0028] The design unit is used to generate prompt templates based on the user's interest graph and the characteristics of the recommended items. The prompt templates include the title, description, selling points of the recommended items, and keywords and phrases related to the user's interests.

[0029] The matching unit is used to match existing items with prompt templates and a large language model to find candidate items that match the user's interests.

[0030] As a preferred option, it also includes an optimization module, used to continuously optimize the recommendation process and prompt templates based on user feedback data and behavioral data.

[0031] This invention deeply mines users' historical interaction data, utilizes the text generation capabilities of large models to capture and represent users' multi-layered interests, and combines designed prompt templates to perform large model retrieval in existing projects to generate explanatory text for recommended items. Attached Figure Description

[0032] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0033] Figure 1 This is a flowchart illustrating the user interest depth mining and multi-level interpretability recommendation method according to an embodiment of the present invention. Detailed Implementation

[0034] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0035] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0036] Example 1:

[0037] like Figure 1 As shown, this embodiment of the invention provides a method for in-depth mining of user interests and multi-level interpretability recommendation, including the following steps:

[0038] The data mining module is used to obtain user interest characteristics based on the user's historical interaction data;

[0039] Step S2: Obtain the user interest map based on user interest characteristics;

[0040] Step S3: Based on the user interest graph and the prompt template, obtain candidate items that match the user's interests;

[0041] Step S4: Based on the candidate items, obtain attractive and explainable recommendation copy;

[0042] Step S5: Continuously optimize the recommendation process and prompt templates based on user feedback data and behavioral data.

[0043] As one implementation of this invention, the user's historical interaction data in step S1 includes browsing history, click behavior, and purchase history. A large language model is used to perform in-depth mining of the user's historical interaction data to obtain the user's explicit and implicit interest features. Distributed prompt design and multi-level in-depth mining are performed using the large language model to obtain the user's comprehensive interest map. The distributed prompt design involves: firstly, introducing the concept of "Promptlets," which are small, specific functional prompts, each focusing on capturing a particular aspect of the user's interest or intent. Then, multiple sets of promptlet combinations (called "Promptlet Ensembles") are designed, each set targeting different scenarios or user groups to achieve more refined interest capture. Explicit interest features can be obtained by analyzing the user's browsing history and click behavior, while implicit interest features are obtained by utilizing the semantic understanding capabilities of the large language model to mine potential interest points from the user's text data.

[0044] As one embodiment of the present invention, step S2 includes:

[0045] Based on user interest characteristics, a multi-level user interest representation model is constructed. This model includes multiple levels of interest nodes, each representing a specific interest or need. A multi-level deep mining approach is employed: Level 1: A large language model is used to perform preliminary analysis of users' historical interaction data, identifying explicit interests and behavioral patterns. Level 2: Promptlet Entries further guide the large language model to uncover users' implicit interests and potential needs, forming an "implicit interest cloud map." Level 3: Combining users' social network information and contextual data, a graph neural network (GNN) is used to construct a comprehensive interest graph, enabling cross-domain and cross-time interest discovery. Simultaneously, an "interest drift detector" is designed to monitor changes in user interests in real time and dynamically adjust the composition and weights of the Promptlet Entries. Utilizing the continuous learning capability of the large language model, new knowledge is continuously extracted from newly added interaction data to update the comprehensive interest graph.

[0046] By connecting interest nodes, a complete user interest graph is formed, which can more comprehensively represent users' interests and needs.

[0047] As one embodiment of the present invention, step S3 includes:

[0048] Based on the user's interest graph and the characteristics of the recommended items, a prompt template is obtained; the prompt template includes the title, description, selling points of the recommended items, as well as keywords and phrases related to the user's interests;

[0049] This approach combines prompt templates with a large language model to perform rapid retrieval and matching within existing projects, finding candidate items that match user interests. Specifically, it involves: first, using a large language model to extract semantic anchors from the user's interest graph. These anchors, as micro-expressions of user interests, accurately point to specific aspects of user preferences; then, dynamically weaving personalized prompt templates based on these semantic anchors. Each time a search is performed, the system re-weaves a customized prompt template based on the user's current interests; finally, using a Rapid Semantic Matching Network (RSMN) within the large language model for rapid retrieval and matching of candidate items. RSMN employs a lightweight network architecture to reduce computational complexity and improve response speed. It typically consists of an input layer, multiple hidden layers, and an output layer. It receives two types of input: feature vectors from the user query and feature vectors from candidate items, which are extracted through a large language model or other pre-trained models. The hidden layers are responsible for transforming the input feature vectors into higher-level semantic representations. These layers use techniques such as three convolutional layers, one pooling layer, one fully connected layer, and a three-channel attention mechanism to capture the complex relationships between the inputs. After processing by the hidden layers, RSMN uses a dedicated matching mechanism to calculate the similarity between the user query and candidate items. This mechanism can be cosine similarity. The output layer is responsible for generating the final matching score, which represents the relevance of the user query to each candidate item. These scores can be normalized to allow for direct comparison between different candidate items. Finally, a list of candidate items that match the user's current interests is output, realizing interest-driven real-time retrieval. The algorithm can efficiently retrieve the candidate items most relevant to the user's current interests from a large item database. This process is real-time and dynamically updated as the user's interests change.

[0050] In one embodiment of the present invention, in step S4, for the retrieved candidate items, attractive and interpretable recommendation copy is generated using the text generation capabilities of a large language model. Specifically: First, the sentiment analysis capabilities of the large language model are used to extract sentiment anchors. These anchors reflect the user's emotional preferences and needs in specific contexts, providing personalized emotional basis for generating recommendation copy. Then, based on the sentiment anchors and the characteristics of the candidate items, a copywriting framework is dynamically constructed. This framework flexibly adapts to the emotional needs of different users, ensuring that the generated copy is both attractive and effectively conveys the value of the items. Next, explanatory elements are incorporated to generate attractive and interpretable recommendation copy. Finally, through real-time feedback and iterative optimization, the quality of the copy and user satisfaction are continuously improved. The copy not only includes the basic information and selling points of the recommended items but also incorporates keywords and phrases related to user interests, thereby enhancing the relevance and attractiveness of the copy. Simultaneously, by structuring the copy, the reasons and basis for the recommendation can be highlighted, improving the interpretability of the recommendation results.

[0051] In one embodiment of the present invention, step S5 involves analyzing user feedback data and behavioral data to identify problems and shortcomings in the recommendation system, allowing for timely adjustments and optimizations. This feedback loop mechanism ensures that the recommendation system remains synchronized with user interests and needs, improving the accuracy and satisfaction of recommendations.

[0052] This invention achieves multi-level interpretable recommendations by deeply mining users' historical interaction data and combining it with the powerful text processing capabilities of large language models. The method of this invention not only captures users' explicit interests but also deeply mines their implicit interests, providing highly personalized and attractive recommendations for each user.

[0053] The following is the implementation process of the method of this invention in the context of movie recommendation, including the following steps:

[0054] Step 1: Data Collection and Preprocessing

[0055] First, historical user interaction data is collected from multiple data sources (such as movie ticketing platforms, social media, and film review websites). Data preprocessing includes data cleaning, word segmentation, and encoding to ensure the accuracy and applicability of the data.

[0056] Step 2: Mining explicit and implicit interest features

[0057] The explicit interest features are mainly extracted from the user's explicit behaviors, as follows:

[0058] Keyword analysis: Extract movie-related keywords from user search history and reviews. Let user u's search history be S. u Comment history is C u The set of explicit keywords is then:

[0059]

[0060] Here, KeywordsExtraction represents the keyword extraction function.

[0061] Content Classification: Movie content is classified using a large language model. Let the set of movies be M, and the set of movie categories be C, then the classification function is:

[0062] (f class :M→C)

[0063] For the set of movies M_u watched by user u, its explicit category interest distribution is as follows:

[0064]

[0065] Here, I is an indicator function that returns 1 if the internal condition is true, and 0 otherwise.

[0066] Behavioral pattern recognition: Analyzing users' viewing time and rating behavior. Let user u rate movie m as r_{u,m} and the viewing time be t_{u,m}. Then, explicit behavioral features can include average rating, rating variance, and viewing time distribution.

[0067] Latent interest characteristics focus on users' potential needs and emotional inclinations, specifically:

[0068] Contextual understanding: Analyzing the semantics and sentiment of user comments. Let user u's comment on movie m be comm_{u,m}, then the implicit sentiment can be obtained using a sentiment analysis model:

[0069] (s u,m =SentimentAnalysis(comm u,m ))

[0070] SentimentAnalysis represents the sentiment analysis function, which returns a sentiment tendency (such as positive, negative, or neutral).

[0071] Topic Modeling: Latent topics are extracted from user comments using topic models such as LDA. Let the topic set be T, and the comment set of user u be C_u. Then the latent topic interest distribution is:

[0072]

[0073] Here, LDA stands for LDA Topic Model, and the return value is the distribution of users across various topics.

[0074] Step 3: Feature Fusion and Representation

[0075] Explicit and implicit interest features are fused into a user interest representation. Let the explicit feature vector be... The latent feature vector is The merged user interests are then represented as follows:

[0076]

[0077] Where α∈[0,1] is a hyperparameter that adjusts the weights of explicit and implicit interests.

[0078] Step 4: Design a prompt template

[0079] In the design of the suggestion template, we can formalize it as a template generation process based on user interest graphs and movie attributes. This process may include selecting appropriate slots, determining the content to fill those slots, and combining this content into a structured recommendation suggestion. The suggestion template generation formula is expressed as follows:

[0080] Given a user interest graph (G_u) and a set of movie attributes (A_m), the generation of the cue template (T) can be represented as:

[0081] [T=f template (G u A m )]

[0082] Among them, f template It is a template generation function that generates personalized recommendation prompt templates based on the user interest graph (G_u) and the movie attribute set (A_m).

[0083] The user interest graph (G_u) can be represented as a graph structure, where nodes represent different points of interest (such as movie genre, director, actors, etc.), and edges represent the relationships between these points of interest. The movie attribute set (A_m) contains various attributes of the movie, such as genre, director, cast list, release date, synopsis, etc.

[0084] function f template The internal logic can be further refined into the following steps:

[0085] Slot selection: Based on the important nodes in the user interest graph (G_u) and the key attributes in the movie attribute set (A_m), select a set of relevant slots S = {s_1, s_2, ..., s_n}.

[0086] Slot Filling: For each selected slot (s_i), determine its filling content. The filling content can be dynamically generated based on node information in the user interest graph (G_u) and the corresponding attributes in the movie attribute set (A_m). For example, if slot (s_i) represents a movie genre, then its filling content can be the movie genre with the highest weight in the user interest graph or genre information from the movie attribute set.

[0087] Template Combination: Combines the populated slot content into a structured recommendation suggestion. This can be achieved through string concatenation, formatting, and other operations to ensure that the generated suggestion has a clear structure and readability.

[0088] The final generated prompt template (T) will be a personalized, structured recommendation message designed to attract the user's attention and provide key information about the recommended movie. It's important to note that this formulaic representation is primarily intended to provide a clear framework; the actual template generation process may involve more complex logic and algorithms.

[0089] Step 5: Generate explainable recommendation copy

[0090] Recommendations are made based on user interest representations and movie attributes. Let the attribute vector of movie m be... User u's interest matching degree with movie m is:

[0091] (Dot product operation)

[0092] The recommendation list consists of N movies that best match the user's interests. For each recommended movie, a recommendation message is generated, containing the movie's basic information and selling points. Let the basic information of movie m be (info_m), and the set of selling points be (selling_points_m), then the recommendation message can be represented as:

[0093] RecommendationText(m) = info m +Highlight(selling p oints m )

[0094] Among them, Highlight means selecting and highlighting the parts of the selling points that are most relevant to the user's interests.

[0095] Step 6: Continuous Optimization and Feedback Loop

[0096] By collecting user feedback data (such as click-through rate, viewing time, ratings, etc.) and behavioral data (such as browsing paths, search keywords, etc.), the recommendation process and suggestion templates are continuously optimized. This data is used to constantly adjust the weights of explicit and implicit interests, optimize the construction of the user interest graph, and improve the design of suggestion templates. This feedback loop mechanism ensures that the recommendation system always keeps pace with changes in user interests and needs.

[0097] Example 2:

[0098] This invention provides a device for in-depth user interest mining and multi-level interpretability recommendation, comprising:

[0099] The data mining module is used to obtain user interest characteristics based on the user's historical interaction data;

[0100] The module is used to generate a user interest graph based on user interest characteristics;

[0101] The retrieval module is used to obtain candidate items that match the user's interests based on the user's interest graph and prompt templates;

[0102] The generation module generates attractive and explainable recommendation text based on candidate items.

[0103] As one embodiment of the present invention, the user's historical interaction data includes: browsing history, click behavior, and purchase history, and the user's interest features include: the user's explicit interest features and implicit interest features.

[0104] As one embodiment of the present invention, the construction module includes:

[0105] The first building unit is used to construct a multi-level user interest representation model based on user interest features; wherein, the user interest representation model includes multiple levels of interest nodes;

[0106] The second building block is used to connect interest nodes to form a user interest graph.

[0107] As one embodiment of the present invention, the retrieval module includes:

[0108] The design unit is used to generate prompt templates based on the user's interest graph and the characteristics of the recommended items. The prompt templates include the title, description, selling points of the recommended items, and keywords and phrases related to the user's interests.

[0109] The matching unit is used to match existing items with prompt templates and a large language model to find candidate items that match the user's interests.

[0110] As one embodiment of the present invention, it further includes: an optimization module, used to continuously optimize the recommendation process and prompt templates based on user feedback data and behavioral data.

[0111] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for user interest deep mining and multi-level explainable recommendation, characterized in that, Includes the following steps: Step S1: Obtain user interest characteristics based on the user's historical interaction data; Step S2: Obtain the user interest map based on user interest characteristics; Step S3: Based on the user interest graph and the prompt template, obtain candidate items that match the user's interests; Step S4: Based on the candidate items, obtain attractive and explainable recommendation copy; In step S1, the user's historical interaction data includes: browsing history, click behavior, and purchase history; the user's interest characteristics include: the user's explicit interest characteristics and implicit interest characteristics. Step S2 includes: Based on user interest features, a multi-level user interest representation model is constructed; the user interest representation model includes multiple levels of interest nodes, each node representing a specific interest point or need. Connect the interest nodes to form a complete user interest graph; Furthermore, in step S2, a multi-level deep mining method is adopted: First level: A large language model is used to perform preliminary analysis of the user's historical interaction data to identify explicit interests and behavioral patterns; Second level: PromptletEnsembles are used to further guide the large language model to mine the user's implicit interests and potential needs, forming an implicit interest cloud map; Third level: Combining the user's social network information and contextual data, a graph neural network is used to construct the user's comprehensive interest map, achieving cross-domain and cross-time interest discovery; Simultaneously, an interest drift detector is designed to monitor changes in user interests in real time and dynamically adjust the composition and weights of the Promptlet Ensembles; Each Promptlet captures a specific aspect of the user's interest or intent; Multiple sets of Promptlet combinations are designed, called Promptlet Ensembles, each set targeting different scenarios or user groups to achieve more refined interest capture; Step S3 includes: Based on the user interest graph and the characteristics of recommended items, a prompt template is obtained. The prompt template includes the title, description, selling points of the recommended items, and keywords and phrases related to user interests. Based on the prompt template and the large language model, a rapid search and matching is performed in the existing projects to find candidate items that match the user's interests. The large language model is used to extract semantic anchors from the user interest graph, and these semantic anchors serve as micro-expressions of user interests. Personalized prompt templates are dynamically woven based on the semantic anchors. It also includes: step S5, continuously optimizing the recommendation process and prompt templates based on user feedback data and behavioral data.

2. A user interest deep mining and multi-level interpretability recommendation device for implementing the user interest deep mining and multi-level interpretability recommendation method of claim 1, characterized in that, include: The data mining module is used to obtain user interest characteristics based on the user's historical interaction data; The module is used to generate a user interest graph based on user interest characteristics; The retrieval module is used to obtain candidate items that match the user's interests based on the user's interest graph and prompt templates; The generation module is used to generate attractive and explainable recommendation text based on candidate items; User historical interaction data includes: browsing history, click behavior, and purchase history; user interest characteristics include: explicit and implicit interest characteristics. The building blocks include: The first building unit is used to construct a multi-level user interest representation model based on user interest features; wherein, the user interest representation model includes multiple levels of interest nodes; The second building unit is used to connect interest nodes to form a user interest graph; The search module includes: The design unit is used to generate prompt templates based on the user's interest graph and the characteristics of the recommended items. The prompt templates include the title, description, selling points of the recommended items, and keywords and phrases related to the user's interests. The matching unit is used to match existing items with prompt templates and a large language model to find candidate items that match the user's interests; It also includes an optimization module, which continuously optimizes the recommendation process and prompt templates based on user feedback and behavioral data.

Images