An e-commerce recommendation method based on large language model for extracting consumption behavior features

Abstract

The application belongs to the field of e-commerce big data recommendation, and particularly relates to an e-commerce recommendation method for extracting consumption behavior features based on a large language model, comprising: obtaining e-commerce platform user and commodity data; forming an external knowledge base through a non-blocked RAG, inputting comments of the user on the commodity, using an LLM in combination with the knowledge base to extract consumption behavior features of the user; calculating final consumption behavior sequence features through fusion of the user behavior sequence and the consumption behavior features; calculating an implicit relationship between user portrait data and target commodity data; inputting the user consumption behavior sequence and the calculated implicit relationship into a multilayer perceptron for prediction to obtain a prediction result. The application can effectively improve the recommendation accuracy of commodities on the e-commerce platform by extracting consumption behavior features of the user through an LLM in combination with a knowledge base, and fusing the features with the behavior sequence.

Description

Technical Field

[0001] This invention belongs to the field of e-commerce big data recommendation, specifically involving an e-commerce recommendation method based on extracting consumer behavior features using a large language model. Background Technology

[0002] With the development of information technology, many domestic and international e-commerce platforms are paying increasing attention to the effectiveness of their product recommendations, focusing on personalized and precise marketing strategies. The accuracy of e-commerce product recommendations not only affects the platform's revenue but also user satisfaction and the overall shopping experience.

[0003] In current e-commerce platforms, operations teams strive to deeply understand users' interests and preferences regarding products. However, traditional technologies struggle to quantify users' emotional responses to recommendation systems and the impact of these recommendations on brand loyalty. To better meet user needs, e-commerce platforms typically rely on user interaction data with recommended products, such as browsing history, purchase behavior, and dwell time, to evaluate the effectiveness of recommendation systems. By analyzing user behavior data, recommendation algorithms can learn user interest patterns and recommend products accordingly. This personalized recommendation not only enhances the user's shopping experience but also increases user stickiness to the platform, thereby improving conversion rates and sales.

[0004] Currently, in the research direction of recommender model construction, scholars have conducted extensive research on content-based recommender models and LLM-based recommender models. The research results are mainly reflected in the following aspects: Firstly, with the development of large language models, researchers can add additional contextual information to large models to achieve input text enhancement. Secondly, the combination of large models and traditional recommender models focuses on the interpretability and transparency of recommendations. In addition, some models emphasize the simplification of recommender models. However, current recommender models still have the following shortcomings:

[0005] 1. Traditional recommendation models ignore rich textual information. Traditional recommendation systems represent user and item features using different embedding IDs. This method learns from the interactions between users and items, effectively capturing potential relationships between them. However, embedding ID-based recommendation systems ignore rich textual information related to users and items. The lack of this additional information leads to less learned content, causing the system to be unable to accurately infer user interests.

[0006] 2. Using large language models as recommenders presents a problem of enormous computational resource consumption. These models typically contain millions or even billions of parameters, which means that their training and inference processes require a large amount of computational resources.

[0007] 3. Large language models suffer from the illusion phenomenon. Large language models rely on massive amounts of text and predict the next word through pattern matching and probability calculation, thereby generating natural language. However, this process lacks real-world logical reasoning capabilities, leading to the possibility that the model may generate inaccurate or even completely false content in complex or ambiguous situations. Summary of the Invention

[0008] To address the aforementioned technical problems, this invention provides an e-commerce recommendation method based on extracting consumer behavior features using a large language model, comprising:

[0009] S1. Obtain user and product data from the e-commerce platform, wherein the user data includes user behavior data and user profile data;

[0010] S2. An external knowledge base is formed through a non-blocked RAG. The input is user reviews of products. The LLM is used in conjunction with the knowledge base to extract the user's consumption behavior characteristics.

[0011] S3. Calculate the final consumption behavior sequence features by fusing user behavior sequences and consumption behavior features;

[0012] S4. Calculate the implicit relationship between user profile data and target product data;

[0013] S5. Input the user consumption behavior sequence and the calculated implicit relationship into the multilayer perceptron for prediction to obtain the prediction result.

[0014] The beneficial effects of this invention are:

[0015] This invention utilizes user reviews, user behavior data, and target product data from e-commerce platforms to target the hidden consumer behavior behind user reviews. It extracts consumer behavior features using LLM combined with a knowledge base, integrates these features with behavior sequences, and updates the consumer behavior representation by updating the behavior update process. By combining the implicit correlations between other features, it predicts product recommendations. This invention can effectively improve the accuracy of product recommendations on e-commerce platforms, achieving precise marketing and recommendation effects. Attached Figure Description

[0016] Figure 1 This is a flowchart of an e-commerce recommendation method based on a large language model for extracting consumer behavior features according to the present invention. Detailed Implementation

[0017] 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.

[0018] An e-commerce recommendation method based on extracting consumer behavior features using a large language model, such as... Figure 1 As shown, it includes:

[0019] S1. Obtain user and product data from the e-commerce platform, wherein the user data includes user behavior data and user profile data;

[0020] S2. An external knowledge base is formed through a non-blocked RAG. The input is user reviews of products. The LLM is used in conjunction with the knowledge base to extract the user's consumption behavior characteristics.

[0021] S3. Calculate the final consumption behavior sequence features by fusing user behavior sequences and consumption behavior features;

[0022] S4. Calculate the implicit relationship between user profile data and target product data;

[0023] S5. Input the user consumption behavior sequence and the calculated implicit relationship into the multilayer perceptron for prediction to obtain the prediction result.

[0024] In this embodiment, sales data and basic data from the e-commerce platform are obtained, including user historical behavior data and product data. The methods for obtaining the data include obtaining raw data from data sources provided by the e-commerce platform or by directly downloading existing publicly available data sources.

[0025] Raw data is usually unstructured and cannot be directly used for data analysis. Simple data cleaning can structure most unstructured data. For example, this involves removing duplicate data and cleaning up invalid nodes, such as some tourist data.

[0026] Preferably, an external knowledge base is formed using a non-blocked RAG (Research Aggregate Object) model. The input is user reviews of products. An LLM (Local Management Model) is used in conjunction with the knowledge base to extract user consumption behavior characteristics, including:

[0027] S21: Assign landmarks (LMKs) to the end of each sentence to achieve special tagging of sentences, and jointly encode the landmarks with the sentences and adjacent contexts to obtain the landmark embedding (LE) representation: LE i ←LLM(c1,...,c i ;LMK).embed[-1];where, the landmark is embedded in LE iDefined as an embedding obtained by jointly encoding landmarks, sentences, and adjacent context, used to represent sentences in a document; c i This represents the sentence in the external document to be passed in, embed[-1] represents the value of the last dimension selected from the output embedding, and LMK represents a special marker after the end of each sentence;

[0028] S22: Obtain the set of user reviews for the product, R, specifically represented as: R = {text...} i |text i ∈U}; where U represents the user set, text i Represents the text of the user's i-th comment;

[0029] S23: Input the set of user reviews of products into the LLM, and let the LLM combine with a knowledge base of consumer behavior theory. The result returned by the LLM is the extracted user consumption behavior characteristics U. c Specifically represented as U c ={u c1 ,u c2 ,…,u ci …,u cn}; where U c This indicates the characteristics of a user's consumption behavior, u ci Let represent the i-th feature of a user, and n represent the total number of features of a user.

[0030] Preferably, by integrating user behavior sequences and behavioral features, attention mechanisms, and stacked autoencoders, a user consumption behavior model and a feature interaction model are established, including:

[0031] S31: Obtain the user behavior sequence feature set U b Specifically, it is represented as: U b ={b1,b2,...,b n}, where b n This represents the user's nth behavior sequence feature;

[0032] S32: Fuse user consumption behavior characteristics and user behavior sequence characteristics to obtain the user consumption behavior sequence set U. bc , represented as: U bc =concat(U b U c )={h′1,h′2,...h′ n}; where U c This represents the user's consumption behavior characteristics; concat() represents the fusion operation; h′ n This represents the user's nth hidden state;

[0033] S33: Based on the user's consumption behavior sequence set U bc Calculate the set of attention scores for each consumer behavior sequence and the target product, denoted as Atns = {α i |i=1,2,...n}; among them, α i This represents the attention score, and n represents the length of the user's consumption behavior sequence.

[0034] α i Attention score is defined as the sequence of each user's consumption behavior, h′. i A similarity measure to the target product q is calculated through a weight parameter allocation mechanism to capture important features in the model.

[0035]

[0036] Where W is the parameter matrix, (h′ i ) T We i This represents the similarity weight between the user's consumption behavior state and the target product q after similarity calculation using a bilinear model.

[0037] S34: Calculate the final consumption behavior sequence features based on the user's consumption behavior sequence feature set and attention score set, expressed as H = AGRY(U bc Atns)={h i |i=1,2,...,n};where, AGRY() represents a gated recurrent unit based on an attention mechanism, h i Let represent the i-th hidden state of the user, and n represent the number of hidden states of the user.

[0038] Considering the dynamic nature of user consumption behavior, the final consumption behavior characteristic is defined as user consumption behavior U. bc Under the attention-based update strategy, the final consumption behavior sequence representation vector is extracted by the user consumption behavior sequence update module.

[0039] Preferably, calculating the implicit relationship between user portrait feature data and target product feature data includes:

[0040] R implicit =SAE(I,P)

[0041] Among them, due to the implicit relationship between features such as target products and user profiles, R is defined. implicit The implicit relationships between these features extracted after feature compression by a stacked encoder are defined, where I is the target product feature set and P is the user portrait feature set.

[0042] Preferably, the user's consumption behavior sequence and the calculated implicit relationship are input into a multilayer perceptron for prediction, and the prediction results include:

[0043] The final consumption behavior sequence features and implicit relationships are concatenated and used as input to a multilayer perceptron:

[0044] [H,R implicit ] = concat(H,R implicit )

[0045] The concatenated features are input into a multilayer perceptron. After several layers of processing, the final predicted interest score is obtained, and the one with the highest interest score is used as the recommendation result.

[0046]

[0047] in, The model predicts interest as a fraction, H represents the final consumption behavior sequence, and R represents the interest score. implicit This indicates the implicit relationship between user profile feature data and target product feature data.

[0048] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An e-commerce recommendation method based on extracting consumer behavior features using a large language model, characterized in that, include: S1. Obtain user and product data from the e-commerce platform, wherein the user data includes user behavior data and user profile data; S2. An external knowledge base is formed through a non-blocked RAG. The input is user reviews of products. The LLM is used in conjunction with the knowledge base to extract the user's consumption behavior characteristics. S21: Assign the landmark LMK to the end of each sentence to achieve special tagging of the sentence, and jointly encode the landmark with the sentence and the adjacent context to obtain the landmark embedding LE representation as follows: Among them, landmark embedding Defined as an embedding obtained by jointly encoding landmarks, sentences, and adjacent context, used to represent sentences in a document; This represents a sentence from the external document to be passed in. This indicates that the value of the last dimension is selected from the output embedding, and LMK represents a special marker after the end of each sentence; S22: Obtain the set R of user reviews for the product, specifically represented as: ;in, Represents a set of users. Indicates the user's first Comment text; S23: Input the set of user reviews of products into the LLM, and combine the LLM with a knowledge base containing the input consumer behavior theory. The result returned by the LLM is the extracted consumer behavior characteristics. Specifically, it is expressed as ;in, This indicates the characteristics of users' consumption behavior. Indicates the user's first One feature; S3. Calculate the final consumption behavior sequence features by fusing user behavior sequences and consumption behavior features; S31: Obtain the user behavior sequence feature set Specifically, it is expressed as: ,in, This represents the user's nth behavior sequence feature; S32: Fuse user consumption behavior characteristics and user behavior sequence characteristics to obtain a set of user consumption behavior sequences. , represented as: ;in, This indicates the characteristics of users' consumption behavior. Indicates a fusion operation. Indicates the user's first One hidden state; S33: Based on the user's consumption behavior sequence set Calculate the set of attention scores for each consumer behavior sequence and the target product, denoted as: ;in, This represents the attention score. Indicates the length of the user's consumption behavior sequence; S34: Based on the user's consumption behavior sequence feature set and attention score set, calculate the final consumption behavior sequence features, expressed as follows: ;in, This represents a gated recurrent unit based on an attention mechanism. Indicates the user's first A hidden state, This indicates the number of hidden states a user has; S4. Calculate the implicit relationship between user profile data and target product data; Calculate the implicit relationship between user profile feature data and target product feature data, including: in, Indicates a stacked automatic encoder. Indicates the characteristics of the target product. Indicates user profile characteristics; S5. Input the final consumption behavior sequence features and the calculated implicit relationships into the multilayer perceptron for prediction to obtain the prediction results; The final consumption behavior u-sequence features and implicit relationships are concatenated and used as the input of a multilayer perceptron; The concatenated features are input into a multilayer perceptron. After several layers of processing, the final predicted interest score is obtained, and the one with the highest interest score is used as the recommendation result.

2. The e-commerce recommendation method based on a large language model for extracting consumer behavior features according to claim 1, characterized in that, The final consumption behavior u-sequence features and implicit relationships are concatenated, including: Where H represents the final consumption behavior sequence, This indicates the implicit relationship between user profile feature data and target product feature data.

3. The e-commerce recommendation method based on a large language model for extracting consumer behavior features according to claim 1, characterized in that, The concatenated features are input into a multilayer perceptron. After several layers of processing, the final predicted interest score is obtained, including: in, H represents the interest score predicted by the model, and H represents the final consumption behavior sequence. This indicates the implicit relationship between user profile feature data and target product feature data.

Images