E-commerce recommendation method for extracting consumption behavior characteristics based on large language model

Abstract

The invention belongs to the field of e-commerce big data recommendation, and particularly relates to an e-commerce recommendation method for extracting consumption behavior characteristics based on a big language model, which comprises the following steps: acquiring e-commerce platform user and commodity data; forming an external knowledge base through RAG without blocks, inputting commodity comments of the user, and extracting consumption behavior characteristics of the user by using LLM in combination with the knowledge base; calculating a final consumption behavior sequence feature by fusing the user behavior sequence and the consumption behavior feature; calculating an implicit relationship between the user portrait data and the target commodity data; and inputting the user consumption behavior sequence and the calculated implicit relationship into a multi-layer perceptron for prediction to obtain a prediction result. According to the method, for the user consumption behaviors hidden behind the user comments, the LLM is combined with the knowledge base to extract the consumption behavior characteristics of the user, and the consumption behavior characteristics are fused with the behavior sequence, so that the commodity recommendation accuracy of the e-commerce platform can be effectively improved.

Description

Technical Field

[0001] The present invention belongs to the field of e-commerce big data recommendation, and specifically relates to an e-commerce recommendation method for extracting consumer behavior characteristics based on a large language model. Background Art

[0002] With the development of information technology, many domestic and foreign Internet e-commerce platforms are paying more and more attention to the recommendation effect of e-commerce platforms, focusing on implementing personalized and precise marketing strategies. The accuracy of e-commerce product recommendations will not only affect the revenue of e-commerce platforms, but also affect user satisfaction and consumption experience.

[0003] In current e-commerce platforms, operations teams are committed to gaining a deep understanding of users’ interests and preferences for products, but it is difficult to quantify users’ emotional responses to recommendation systems and the impact of these recommendations on company brand loyalty using traditional techniques. In order to better meet user needs, e-commerce platforms usually 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 users’ interest patterns and recommend products accordingly. This personalized recommendation not only improves users’ shopping experience, but also increases users’ stickiness to the platform, thereby increasing conversion rates and sales.

[0004] At present, in the research direction of recommendation model construction, scholars have conducted extensive research on content-based recommendation models and LLM-based recommendation models. The research results are mainly reflected in the following aspects: On the one hand, with the development of large language models, researchers can add additional contextual information through large models to achieve input text enhancement. On the other hand, the combination of large models and traditional recommendation models focuses on the interpretability and transparency of recommendations. In addition, there are models that focus on the simplification of recommendation models. However, the current recommendation models still have the following shortcomings:

[0005] 1. Traditional recommendation models ignore rich text information. Traditional recommendation systems use different embedding IDs to represent user and item features. This method learns based on the interaction between users and items and can effectively capture the potential relationship between users and items. However, the recommendation system based on embedded IDs ignores the rich text information related to users and items. The lack of this additional information will lead to less content being learned, which will cause the system to be unable to accurately infer user interests.

[0006] 2. Using large language models as recommenders consumes a lot of computing resources. Models usually contain millions or even billions of parameters, which requires a lot of computing resources for their training and inference processes.

[0007] 3. Large language models have hallucinations. Large language models use massive amounts of text as a basis to predict the next word through pattern matching and probability calculation, thereby achieving natural language generation. However, this process lacks the logical reasoning ability of the real world, resulting in the model generating inaccurate or even completely false content in situations where the information is complex or ambiguous. Summary of the invention

[0008] In order to solve the above technical problems, the present invention provides an e-commerce recommendation method for extracting consumer behavior characteristics based on a large language model, comprising:

[0009] S1. Obtaining user and product data of the e-commerce platform, wherein the user data includes user behavior data and user portrait data;

[0010] S2, form an external knowledge base through RAG without block division, input is user comments on products, use LLM to combine with knowledge base, and extract user consumption behavior characteristics;

[0011] S3, calculating the final consumption behavior sequence characteristics by integrating the user behavior sequence and consumption behavior characteristics;

[0012] S4, calculating the implicit relationship between the user portrait data and the target product data;

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

[0014] Beneficial effects of the present invention:

[0015] The present invention utilizes user comments, user behavior data and target product data of the e-commerce platform, targets the user consumption behavior hidden behind the user comments, extracts the user's consumption behavior characteristics through LLM combined with the knowledge base, merges it with the behavior sequence, obtains the consumption behavior representation by updating its consumption behavior update process, and predicts product recommendations in combination with the invisible associations between other features. The present invention can effectively improve the recommendation accuracy of products on the e-commerce platform and achieve the effect of precision marketing and recommendation. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 The present invention is a flowchart of an e-commerce recommendation method for extracting consumer behavior features based on a large language model. DETAILED DESCRIPTION

[0017] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0018] An e-commerce recommendation method based on a large language model to extract consumer behavior features, such as Figure 1 As shown, including:

[0019] S1. Obtaining user and product data of the e-commerce platform, wherein the user data includes user behavior data and user portrait data;

[0020] S2, form an external knowledge base through RAG without block division, input is user comments on products, use LLM to combine with knowledge base, and extract user consumption behavior characteristics;

[0021] S3, calculating the final consumption behavior sequence characteristics by integrating the user behavior sequence and consumption behavior characteristics;

[0022] S4, calculating the implicit relationship between the user portrait data and the target product data;

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

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

[0025] The raw data usually obtained is unstructured and cannot be used directly for data analysis. Most unstructured data can be structured through simple data cleaning. For example, duplicate data can be deleted and invalid nodes such as some visitor data can be cleaned up.

[0026] Preferably, an external knowledge base is formed through non-blocked RAG, and the input is user comments on products. LLM is used in combination with the knowledge base to extract the user's consumption behavior characteristics, including:

[0027] 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 adjacent context to obtain the landmark embedding LE represented as: LE i ←LLM(c1,...,c i ;LMK).embed[-1]; where the landmark is embedded in LE iIt is defined as the embedding obtained by jointly encoding the landmark with the sentence and the adjacent context, which is used to represent the sentence in the document; c i Represents the sentence in the external document to be passed in, embed[-1] represents selecting the value of the last dimension from the output embedding, and LMK represents a special token at the end of each sentence;

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

[0029] S23: Input the user's review set on the product into LLM, so that LLM combines with the knowledge base of consumer behavior theory knowledge, and the result returned by LLM is the extracted user consumption behavior feature U c , specifically expressed as U c = {u c1 ,u c2 ,…,u ci …,u cn}; Among them, U c represents the consumption behavior characteristics of users, u ci represents the i-th feature of the user, and n represents the total number of features of the user.

[0030] Preferably, a user consumption behavior model and a feature interaction model are established by integrating user behavior sequences and behavior features, an attention mechanism, and a stacked autoencoder, including:

[0031] S31: Obtaining user behavior sequence feature set U b , specifically expressed as: U b ={b1,b2,...,b n}, where b n Indicates the user's nth behavior sequence feature;

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

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

[0034] α i The attention score is defined as each user consumption behavior sequence h′ i A similarity measure with the target item q, calculated through a weight parameter assignment mechanism to capture important features in the model.

[0035]

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

[0037] S34: Calculate the final consumption behavior sequence feature according to 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 the gated recurrent unit based on the attention mechanism, h i represents the i-th hidden state of the user, and n represents the number of hidden states of the user.

[0038] Considering the dynamics of user consumption behavior, the final consumption behavior characteristics are defined as user consumption behavior U bc Under the attention mechanism-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 the user portrait feature data and the target product feature data includes:

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

[0041] Among them, since there is an implicit relationship between the target product and the user portrait and other features, we define R implicit It is the invisible relationship between these features extracted after feature compression by the stacked encoder, where I is the target product feature set and P is the user portrait feature set.

[0042] Preferably, the user consumption behavior sequence and the calculated implicit relationship are input into a multi-layer perceptron for prediction to obtain a prediction result, including:

[0043] The final consumption behavior u sequence features and implicit relationships are concatenated as the input of the multi-layer perceptron:

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

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

[0046]

[0047] in, represents the interest score predicted by the model, H represents the final consumption behavior sequence, and R implicit Represents the implicit relationship between user portrait feature data and target product feature data.

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

Claims

1. An e-commerce recommendation method based on a large language model to extract consumer behavior features, characterized in that: include: S1. Obtaining user and product data of the e-commerce platform, wherein the user data includes user behavior data and user portrait data; S2, form an external knowledge base through RAG without block division, input is user comments on products, use LLM to combine with knowledge base, and extract user consumption behavior characteristics; S3, calculating the final consumption behavior sequence characteristics by integrating the user behavior sequence and consumption behavior characteristics; S4, calculating the implicit relationship between the user portrait data and the target product data; S5. Input the user consumption behavior sequence and the calculated implicit relationship into a multi-layer perceptron for prediction to obtain the prediction result.

2. According to claim 1, an e-commerce recommendation method for extracting consumer behavior features based on a large language model is characterized in that: The external knowledge base is formed by RAG without blocks. The input is the user's comments on the product. LLM is used to combine the knowledge base to extract the user's consumption behavior characteristics, including: 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 adjacent context to obtain the landmark embedding LE represented as: LE i ←LLM(c1,...,c i ;LMK).embed[-1]; where the landmark is embedded in LE i It is defined as the embedding obtained by jointly encoding the landmark with the sentence and the adjacent context, which is used to represent the sentence in the document; c i Represents the sentence in the external document to be passed in, embed[-1] represents selecting the value of the last dimension from the output embedding, and LMK represents a special token after the end of each sentence; S22: Obtain the user's comment set R on the product, specifically expressed as: R = {text i |text i ∈U}; where U represents the user set, text i represents the i-th comment text of the user; S23: Input the user's review set on the product into LLM, so that LLM combines with the knowledge base of consumer behavior theory knowledge, and the result returned by LLM is the extracted user consumption behavior feature U c , specifically expressed as U c = {u c1 ,u c2 ,…,u ci …,u cn }; Among them, U c represents the consumption behavior characteristics of users, u ci represents the i-th feature of the user, and n represents the total number of features of the user.

3. According to claim 1, the e-commerce recommendation method for extracting consumer behavior features based on a large language model is characterized in that: By integrating user behavior sequences and behavior features, attention mechanisms, and stacked autoencoders, we build a user consumption behavior model and feature interaction model, including: S31: Obtaining user behavior sequence feature set U b , specifically expressed as: U b ={b1,b2,...,b n }, where b n Indicates the user's nth behavior sequence feature; S32: Fusing the user's consumption behavior characteristics and the user's behavior sequence characteristics to obtain the user's consumption behavior sequence set U bc , expressed as: U bc =concat(U b ,U c )={h′1,h′2,...h′ n }; Among them, U c represents the consumption behavior characteristics of users, concat() represents the fusion operation, h′ n represents the nth hidden state of the user; S33: Based on the user's consumption behavior sequence set U bc , calculate the attention score set of each consumer behavior sequence and the target product, expressed as Atns = {α i |i=1,2,...n}; among them, α i represents the attention score, and n represents the length of the user's consumption behavior sequence; S34: Calculate the final consumption behavior sequence feature according to the user's consumption behavior sequence feature set and attention score set, expressed as H = AGRU (U bc ,Atns)={h i |i=1,2,...,n}; where AGRU() represents the gated recurrent unit based on the attention mechanism, h i represents the i-th hidden state of the user, and n represents the number of hidden states of the user.

4. The e-commerce recommendation method based on extracting consumer behavior features based on a large language model according to claim 1, characterized in that: Calculate the implicit relationship between the user portrait feature data and the target product feature data, including: R implicit =SAE(I,P) Among them, SAE() represents stacked autoencoder, I represents target product features, and P represents user portrait features.

5. The e-commerce recommendation method based on extracting consumer behavior features based on a large language model according to claim 1, characterized in that: The user consumption behavior sequence and the implicit relationship of the calculation are input into the multi-layer perceptron for prediction, and the prediction results are obtained, including: The final consumption behavior u sequence features and implicit relationships are concatenated as the input of the multi-layer perceptron; The concatenated features are input into a multi-layer perceptron. After several layers of processing, the final predicted interest score is obtained, and the one with the highest interest score is taken as the recommendation result.

6. The e-commerce recommendation method based on extracting consumer behavior features based on a large language model according to claim 5, characterized in that: The final consumption behavior u sequence features and implicit relationships are spliced ​​together, including: [H,R implicit ]=concat(H,R implicit ) Among them, H represents the final consumption behavior sequence, R implicit Represents the implicit relationship between user portrait feature data and target product feature data.

7. The e-commerce recommendation method based on extracting consumer behavior features based on a large language model according to claim 5, characterized in that: The concatenated features are input into a multi-layer perceptron. After several layers of processing, the final predicted interest score is obtained, including: in, represents the interest score predicted by the model, H represents the final consumption behavior sequence, and R implicit Represents the implicit relationship between user portrait feature data and target product feature data.

Images