Consumption decision intention analysis method and device for e-commerce platform

Abstract

The invention provides a consumption decision intention analysis method and device for an e-commerce platform, and the method comprises the steps: generating an initial prompt text through a prompt text generator based on a search word, and user group data, category data and / or time-space field data; performing vector retrieval in an external RAG knowledge base based on the search word, and splicing a retrieval result into the initial prompt text to obtain a target prompt text; inputting the target prompt text into a pre-trained e-commerce vertical domain LLM to obtain a decision factor set representing a consumption intention decision; and performing trend identification on the decision factor set according to a pre-trained consumption trend identification LLM to obtain a concentration description result of the decision factors. The consumption decision analysis accuracy can be improved.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and particularly to a method and device for analyzing consumption decision-making intentions on an e-commerce platform. Background Art

[0002] On an e-commerce platform, there are many decision-making factors that affect consumers' orders. These factors can be internal or external, and they jointly act on the consumers' thinking process, prompting consumers to form specific decisions. Due to the numerous decision-making factors, it is difficult to analyze and understand users' preferences, and to determine which factors play a decisive role and their importance in specific scenarios. Therefore, for those skilled in the art, how to effectively analyze consumption decision-making intentions is an important technical problem. Summary of the Invention

[0003] In view of this, this application provides a method and device for analyzing consumption decision-making intentions on an e-commerce platform, with the main purpose of effectively analyzing consumption decision-making intentions on the e-commerce platform.

[0004] According to one aspect of this application, a method for analyzing consumption decision-making intentions on an e-commerce platform is provided, including:

[0005] Based on the search term, as well as user population data, category data, and / or spatio-temporal field data, use a prompt text generator to generate an initial prompt text;

[0006] Perform vector retrieval on the external RAG knowledge base based on the search term, and splice the retrieval results into the initial prompt text to obtain a target prompt text;

[0007] Input the target prompt text into a pre-trained e-commerce vertical domain LLM to obtain a set of decision factors representing consumption intention decisions, where the e-commerce vertical domain LLM is obtained by continuously pre-training a general LLM based on e-commerce vertical domain data;

[0008] According to a pre-trained consumption trend recognition LLM, perform trend recognition on the set of decision factors to obtain a concentration characterization result of the decision factors, where the consumption trend recognition LLM is obtained by training and reasoning on the e-commerce vertical domain LLM based on serialized samples including user decision attributes, category attributes, and / or spatio-temporal attributes constructed from user population data, category data, and / or spatio-temporal field data.

[0009] In one implementation, the performing vector retrieval on the external RAG knowledge base based on the search term, and splicing the retrieval results into the initial prompt text to obtain a target prompt text includes:

[0010] Receive the search term and perform regularization, rewriting, and query expansion processing on the search term;

[0011] Perform vector retrieval in an external RAG knowledge base based on the result after processing the search term, and perform chain-of-thought re-ranking on the retrieval results;

[0012] Inject the retrieval results into the initial prompt text in the re-ranking order to obtain the target prompt text.

[0013] In one implementation, the method further includes:

[0014] Construct an external RAG knowledge base, where one or more dimensional data of personal factors, product factors, scenario factors, information communication factors, and technology and interface factors are statistically analyzed, and a set of decision factors for consumption decisions is determined based on the dimensional data to form the external RAG knowledge base,

[0015] Among them, the personal factors include demand / desire factors, economic status factors, psychological factors, and cognitive factors; the product factors include quality factors, price factors, functional factors, and design and appearance factors; the scenario factors include shopping occasion factors and social and cultural factors; the information communication factors include evaluation and review factors, advertising and promotion factors, trust and security factors, and after-sales service factors; the social and network factors include word-of-mouth communication factors, community influence factors, and network effect factors; the technology and interface factors include user experience factors, mobile compatibility factors, and technical support factors.

[0016] In one implementation, continuously pre-train the general LLM based on e-commerce vertical domain data, including:

[0017] Obtain e-commerce vertical domain data from the material library within the e-commerce vertical domain, the knowledge graph within the vertical domain, and / or the image library within the vertical domain;

[0018] Extract features from the e-commerce vertical domain data to determine the data expressions of the knowledge category dimension and the product dimension. The data expressions of the knowledge category dimension include any one or more of attribute information, category information, and label information, and the data expressions of the product dimension include any one or more of brand, taste, merchant, and evaluation;

[0019] Determine training samples based on the data expressions of the knowledge category dimension and the product dimension of the e-commerce vertical domain data, and continuously pre-train the general LLM.

[0020] In one implementation, it further includes:

[0021] Perform reward modeling on the e-commerce vertical domain LLM, where a series of text responses are generated from the e-commerce vertical domain LLM, scored or classified based on the text responses to form feedback data; based on the collected feedback data, train a reward model; use a reinforcement learning algorithm, take the e-commerce vertical domain LLM as an agent, and fine-tune the e-commerce vertical domain LLM through interaction with the reward model.

[0022] In one implementation, the construction process of the consumption trend recognition LLM includes:

[0023] Obtain user population data, including personal information of users, consumption habit data, and preference data; obtain category data, including classification data, feature data, and price data of goods or services; obtain spatio-temporal field data, including timestamp information and geographical location information;

[0024] Based on the categories of goods or services, form a certain number of serialized sub-elements for each category, and aggregate the serialized resources as the serialized samples of the category;

[0025] Based on the model architecture of generative recommendation, use a hierarchical sequence conversion unit structure to process the serialized samples, and fine-tune the e-commerce vertical domain LLM according to the processing results to obtain a consumption trend recognition LLM with sequence modeling.

[0026] In one implementation, it further includes:

[0027] Based on a dual-model self-supervised mechanism, control the e-commerce vertical domain LLM and the consumption trend recognition LLM to supervise each other.

[0028] In one implementation, it further includes:

[0029] Apply the concentration characterization result of the decision factor to any of the following scenarios: search scenario, optimize the search result ranking based on the concentration characterization result of the decision factor; product selection scenario, optimize the product category selection or adjust the combination strategy of existing products based on the concentration characterization result of the decision factor; recommendation scenario, push goods or services with high interest to consumers based on the concentration characterization result of the decision factor; advertising scenario, optimize the advertising placement strategy based on the concentration characterization result of the decision factor; operation scenario, provide an overall operation strategy for merchants based on the concentration characterization result of the decision factor.

[0030] According to one aspect of the present application, a modeling method for analyzing consumption decision intentions of an e-commerce platform includes:

[0031] Based on e-commerce vertical domain data, continuously pre-train a general LLM to obtain an e-commerce vertical domain LLM;

[0032] Construct serialized samples including user decision attributes, category attributes, and / or spatio-temporal attributes based on user population data, category data, and / or spatio-temporal field data, and train and infer the e-commerce vertical LLM based on the serialized samples to obtain a consumption trend recognition LLM;

[0033] Among them, when analyzing the consumption decision intention, first determine the target prompt text and use the target prompt text as the input of the e-commerce vertical LLM to obtain a set of decision factors representing the consumption intention decision; then, perform trend prediction on the set of decision factors according to the consumption trend recognition LLM to obtain the concentration characterization result of the decision factors; the process of determining the target prompt text includes: generating an initial prompt text based on the search term, as well as user population data, category data, and / or spatio-temporal field data, using a prompt text generator; performing vector retrieval on the external RAG knowledge base based on the search term, and splicing the retrieval result into the initial prompt text to obtain the target prompt text.

[0034] In one implementation, continuous pre-training of the general LLM based on e-commerce vertical data includes:

[0035] Obtain e-commerce vertical data from the material library within the e-commerce vertical domain, the knowledge graph within the vertical domain, and / or the picture library within the vertical domain;

[0036] Extract features from the e-commerce vertical data to determine the data expressions of the knowledge class dimension and the product dimension. The data expressions of the knowledge class dimension include any one or more of attribute information, category information, and label information, and the data expressions of the product dimension include any one or more of brand, flavor, merchant, and evaluation;

[0037] Based on the data expressions of the knowledge class dimension and the product dimension of the e-commerce vertical data, determine training samples and perform continuous pre-training on the general LLM.

[0038] In one implementation, it further includes:

[0039] Perform reward modeling on the e-commerce vertical LLM. Among them, generate a series of text responses from the e-commerce vertical LLM, score or classify based on the text responses to form feedback data; train a reward model based on the collected feedback data; use the reinforcement learning algorithm, take the e-commerce vertical LLM as an agent, and fine-tune the e-commerce vertical LLM through interaction with the reward model.

[0040] In one implementation, the construction process of the consumption trend recognition LLM includes:

[0041] Obtain user population data, including users' personal information, consumption habit data, and preference data; obtain category data, including classification data, feature data, and price data of goods or services; obtain spatio-temporal field data, including timestamp information and geographical location information;

[0042] Based on the category of goods or services, a certain number of serialized sub-elements of each category are formed, and the serialized resources are aggregated as the serialized sample of the category;

[0043] Based on the model architecture of generative recommendation, a hierarchical sequence conversion unit structure is used to process the serialized sample, and the e-commerce vertical LLM is fine-tuned according to the processing result to obtain a consumption trend recognition LLM for sequence modeling.

[0044] In one implementation, it further includes:

[0045] Based on the dual-model self-supervised mechanism, control the e-commerce vertical LLM and the consumption trend recognition LLM to supervise each other.

[0046] In one implementation, the vector retrieval in the external RAG knowledge base based on the search term, and splicing the retrieval result into the initial prompt text to obtain the target prompt text, includes:

[0047] Receive the search term, and perform regularization, rewriting, and query expansion processing on the search term;

[0048] Perform vector retrieval in the external RAG knowledge base based on the result of the processed search term, and perform thought-chain re-ranking on the retrieval result;

[0049] Inject the retrieval result into the initial prompt text in the re-ranking order to obtain the target prompt text.

[0050] In one implementation, the method further includes:

[0051] Construct an external RAG knowledge base, where one or more dimensional data of personal factors, product factors, scenario factors, information communication factors, and technology and interface factors are statistically analyzed, and a decision factor set for consumption decisions is determined based on the dimensional data to form the external RAG knowledge base,

[0052] Among them, the personal factors include demand / desire factors, economic status factors, psychological factors, and cognitive factors; the product factors include quality factors, price factors, functionality factors, design and appearance factors; the scenario factors include shopping occasion factors, social and cultural factors; the information communication factors include evaluation and review factors, advertising and promotion factors, trust and security factors, and after-sales service factors; the social and network factors include word-of-mouth communication factors, community influence factors, and network effect factors; the technology and interface factors include user experience factors, mobile compatibility factors, and technical support factors.

[0053] In one implementation, it further includes:

[0054] Apply the concentration characterization result of the decision factor to any of the following scenarios: search scenario, optimize the search result sorting based on the concentration characterization result of the decision factor; product selection scenario, optimize the product category selection or adjust the combination strategy of existing products based on the concentration characterization result of the decision factor; recommendation scenario, push products or services with high interest to consumers based on the concentration characterization result of the decision factor; advertising scenario, optimize the advertising placement strategy based on the concentration characterization result of the decision factor; operation scenario, provide an overall operation strategy for merchants based on the concentration characterization result of the decision factor.

[0055] According to one aspect of the present application, a device is provided for executing the above method.

[0056] According to one aspect of the present application, a storage medium is provided, in which a computer program is stored, and the computer program is set to execute the above method when running.

[0057] According to one aspect of the present application, an electronic device is provided, including a memory and a processor. A computer program is stored in the memory, and the processor is set to run the computer program to execute the above method.

[0058] By means of the above technical solution, the method for analyzing the consumption decision intention of the e-commerce platform provided by the present application is different from the conventional method of training a general LLM using vertical domain data to obtain a vertical domain LLM. Instead of simply training a general LLM using e-commerce vertical domain data to obtain an e-commerce vertical domain LLM, in the process of using the e-commerce vertical domain LLM for intention recognition, in addition to relying on the internal knowledge base of the LLM, an external RAG knowledge base is also combined, so as to incorporate a large number of pre-statistically summarized decision factors, thereby realizing retrieval enhancement and improving the accuracy and efficiency of consumption decision recognition. In addition, based on the e-commerce vertical domain LLM, a consumption trend recognition LLM is further trained, and the consumption trend recognition LLM is used to perform trend inference on the recognition results of the vertical domain LLM, so as to give the concentration characterization result of the consumption decision intention.

[0059] In summary, the solution of this application has at least the following remarkable technical effects:

[0060] First, not only train the e-commerce vertical domain LLM, but also further train the consumption trend recognition LLM based on the e-commerce vertical domain LLM. That is, use a dual-model to identify the trend of consumption intention. The dual-models are self-supervised (independent and mutually supervised with each other), which can effectively improve the recognition accuracy;

[0061] Second, in the process of using the e-commerce vertical domain LLM to identify decision factors, not only based on the internal knowledge base of the model, but also adopt the RAG method to use the external RAG knowledge base for retrieval enhancement, expanding the retrieval scope to improve the retrieval accuracy;

[0062] Third, adopt a two-step recognition mode. First, use the e-commerce vertical domain LLM to make a "decision" to determine the set of decision factors for the current search term (Query). Then, use the consumption trend recognition LLM to give the concentration characterization result of the set of decision factors, realizing the prediction of the consumption decision trend. Compared with the conventional method that only gives the recognition result without trend analysis, the consumption decision trend analysis of this application has guiding significance for the specific scenarios (product selection, recommendation, advertising, operation, etc.) of the e-commerce platform.

[0063] The above description is only an overview of the technical solution of this application. In order to be able to understand the technical means of this application more clearly, it can be implemented according to the content of the specification. And in order to make the above and other purposes, features and advantages of this application more obvious and understandable, the following specifically gives the specific implementation manners of this application. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] The drawings described herein are used to provide a further understanding of this application and form a part of this application. The illustrative embodiments of this application and their descriptions are used to explain this application and do not constitute an improper limitation of this application. In the drawings:

[0065] Figure 1 Shows a schematic diagram of a modeling method for analyzing consumption decision intention of an e-commerce platform provided by an embodiment of this application;

[0066] Figure 2 Shows a schematic diagram of a method for analyzing consumption decision intention of an e-commerce platform provided by an embodiment of this application;

[0067] Figure 3 Shows a schematic diagram of data expression in the product dimension provided by an embodiment of this application;

[0068] Figure 4 Shows a schematic diagram of the domain knowledge base and the external knowledge base provided by an embodiment of this application;

[0069] Figure 5 Shows the application schematic diagram of RAG in the vertical domain LLM training stage of the embodiments of the present application;

[0070] Figure 6 Shows the application schematic diagram of RAG in the question-and-answer stage of the embodiments of the present application;

[0071] Figure 7 Shows the schematic diagram of the dual-model of the embodiments of the present application for analyzing the consumption decision-making intention. Detailed implementation manners

[0072] In order to enable those skilled in the art to better understand the solutions of the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present application. It should be noted that, without conflict, the embodiments in the present application and the features in the embodiments may be combined with each other.

[0073] The inventors of the present application have found through research that there are many factors (decision factors) that affect the decision-making of consumers on e-commerce platforms before placing an order. In e-commerce trading platforms, "decision factors" refer to various factors that directly affect or influence consumers to make purchase decisions. These factors can be internal or external, and they jointly act on the thinking process of consumers, prompting consumers to form specific decisions. Understanding decision factors is crucial for e-commerce platforms and merchants, as it helps them design more effective marketing strategies, optimize the user experience, and improve the sales conversion rate.

[0074] Due to the numerous decision factors, it is relatively difficult to analyze and understand the preferences of users, and in specific scenarios, which factors play a decisive role and what is their importance. This affects scenario optimization and algorithm model optimization, and also causes a certain degree of inexplicability. Therefore, the present application proposes to rely on large language models to gain insights into and explain consumption decision-making intentions.

[0075] Large Language Model (LLM) is an important research direction and technological breakthrough in the field of Natural Language Processing (NLP) in recent years. The core of large language models lies in their scale and complexity. They usually have billions or even trillions of parameters, much larger than previous models, which enables them to process and understand extremely complex language structures and semantics.

[0076] With the development of large language models (LLMs), the models have demonstrated powerful capabilities. However, the computing power required for model training is also increasing. To reduce the training cost of the models, domain-specific LLMs have received increasing attention. A domain-specific LLM (vertical domain LLM) refers to a model that, in a specific domain (such as e-commerce, finance, education, medicine, code, mathematics, etc.), is trained or fine-tuned using industry data on a general LLM to meet the professional knowledge and skill requirements of the specific domain. The main difference between a domain-specific LLM and a general LLM lies in its application scenario and professional knowledge. A general LLM is pre-trained using a large amount of general text data and has cross-task and cross-domain generality. In contrast, a domain-specific LLM focuses on a specific domain and is trained or fine-tuned using industry data to provide more professional and practical services. The application of domain-specific LLMs in specific domains can significantly improve production efficiency and reduce production costs.

[0077] In this application, a domain-specific e-commerce LLM is trained, and based on this, the consumer decision-making intention is analyzed. Different from the conventional method of training a general LLM using domain-specific data to obtain a domain-specific LLM, the domain-specific e-commerce LLM in this application not only uses e-commerce domain-specific data (such as in-domain material libraries, in-domain knowledge graphs, in-domain image libraries) to train the general LLM to obtain the domain-specific e-commerce LLM, but also, during the process of using the domain-specific e-commerce LLM for intention recognition, in addition to relying on the internal knowledge base of the LLM, combines an external RAG knowledge base to incorporate a large number of pre-statistically aggregated decision factors, thereby achieving retrieval enhancement and improving the accuracy and efficiency of consumer decision-making recognition. In addition, based on the domain-specific e-commerce LLM, further training (constructing serialized samples including user decision-making attributes, category attributes, and / or spatio-temporal attributes according to user population data, category data, and / or spatio-temporal field data, and training and reasoning on the domain-specific e-commerce LLM based on the serialized samples, which will be described in detail below) is performed to obtain a consumer trend recognition LLM, and the consumer trend recognition LLM is used to perform trend reasoning on the recognition results of the domain-specific LLM, thereby giving a concentration characterization result of the consumer decision-making intention.

[0078] See Figure 1 , which shows a schematic diagram of a modeling method for analyzing consumer decision-making intention on an e-commerce platform provided by an embodiment of this application, and, see Figure 2 , which shows a schematic diagram of a method for analyzing consumer decision-making intention on an e-commerce platform provided by an embodiment of this application. That is, Figure 1 describes the modeling process of the LLM involved in the embodiments of this application, Figure 2 shows the process of using the LLM for analyzing consumer decision-making intention.

[0079] In the embodiments of the present application, the LLM modeling stage includes the following steps: First, based on e-commerce vertical domain data, continuously pre-train (CPT) a general LLM to obtain an e-commerce vertical domain LLM; then, based on user population data, category data, and / or spatio-temporal field data, construct serialized samples including user decision attributes, category attributes, and / or spatio-temporal attributes, and based on the serialized samples, train and infer the e-commerce vertical domain LLM to obtain a consumption trend recognition LLM.

[0080] In the embodiments of the present application, the stage of identifying consumption decision intentions based on the LLM includes the following steps: First, determine a target prompt and use the target prompt as the input of the e-commerce vertical domain LLM to obtain a set of decision factors representing consumption intention decisions; then, perform trend prediction on the set of decision factors according to the consumption trend recognition LLM to obtain a concentration characterization result of the decision factors; among them, the process of determining the target prompt includes: based on the Query, as well as user population data, category data, and / or spatio-temporal field data, use a prompt text generator to generate an initial prompt; perform vector retrieval on the external RAG knowledge base based on the Query, and splice the retrieval results into the initial prompt to obtain the target prompt.

[0081] It can be seen that in order to implement the embodiments of the present application, some data preparation and processing work needs to be done, including the collection and processing of e-commerce vertical domain data, the serialization processing of data, and the construction of an external RAG knowledge base. Therefore, in the modeling stage, use e-commerce vertical domain data to perform CPT on a general LLM to obtain an e-commerce vertical domain LLM, and use serialized samples to train the e-commerce vertical domain LLM to obtain a consumption trend recognition LLM. In addition, in the recognition stage, use external RAG for retrieval enhancement and splice the target prompt.

[0082] The following provides a detailed exemplary description of each key step / technical point of the embodiments of the present application.

[0083] I. Collection and Processing of E-commerce Vertical Domain Data

[0084] It can be understood that the e-commerce mentioned in the embodiments of the present application refers to a specific type of e-commerce platform, such as a certain takeaway APP. In order to construct a vertical domain model for this specific platform, it is necessary to pre-collect and analyze an industry-specific knowledge base. The industry-specific knowledge base includes a knowledge base, an image library, and a material library within the vertical domain. These knowledge bases provide rich background information and context, supporting the model to generate more accurate and targeted text. Taking the takeaway e-commerce platform as an example, through the analysis of the vertical domain knowledge base, the following two-dimensional data expressions can be obtained: the data expression of the knowledge class dimension, and the data expression of the product expression dimension.

[0085] In the embodiments of the present application, after obtaining e-commerce vertical domain data from the material library, knowledge graph, and picture library within the e-commerce vertical domain, feature extraction is performed on the e-commerce vertical domain data to determine the data representation of the knowledge class dimension (including attribute information, category information, tag information, etc.) and the data representation of the product dimension (including attribute information such as brand, taste, merchant, evaluation, etc.).

[0086] The data representation of the knowledge class data dimension includes attribute information, category information, and tag information. As shown in the example of Table 1, it shows the attribute knowledge of some retail goods.

[0087] Table 1

[0088]

[0089] The data representation of the product dimension includes attribute information such as brand, taste, merchant, evaluation, etc. Figure 3 For example, for the commodity of instant noodles, its product representation includes information such as Laotan Instant Noodles, Korean Instant Noodles, Heweida Instant Noodles, Tongyi, Tangdaren, Shin Ramyun, Master Kong, Spicy Fire Chicken Noodles, Quality One-Person Food, Xiaolanxin Merchant, Afternoon Tea Hot Search, and Numerous Positive Reviews.

[0090] II. Perform CPT on the general LLP to obtain the e-commerce vertical domain LLM

[0091] After performing the data representation of the knowledge class dimension and the product dimension on the e-commerce vertical domain data, using the processed e-commerce vertical domain data as training samples, perform CPT (Continued Pre-training) on the general LLP.

[0092] CPT (Compressed Pre-trained Transformer) continues to perform pre-training on the pre-trained LLM on a corpus in a specific domain, enabling the model to better understand data in a specific domain or type. In the embodiments of the present application, distill the general LLM through CPT. The distillation process transfers the knowledge of the general LLM to a smaller and more efficient model, reducing the consumption of computing resources while maintaining the quality of the generated text.

[0093] In the implementation of performing CPT on the general LLM, several aspects need to be noted. 1. Data selection and mixing: During the continued pre-training process, select an appropriate data subset for training and use a high-quality sub-corpus for pre-training. In addition, use mixed data to reduce the distribution gap and improve the adaptability and stability of the model. 2. Training strategy: Adopt multiple strategies to optimize the training process. For example, perform continued pre-training on an appropriate-sized subset for multiple cycles, only perform pre-training on high-quality sub-corpora, and use mixed data similar to the pre-training data to reduce the distribution gap.

[0094] In the embodiments of the present application, during the CPT process of the general LLP, reward modeling can also be implemented according to the Reward mechanism to improve the effective learning performance of the model.

[0095] Reward Modeling is mainly used to train the agent (large language model) on how to learn and follow human-expected behaviors more effectively. In a reinforcement learning environment, the agent tries different behaviors to obtain the reward signals given by the environment, and adjusts its behavior strategy accordingly to maximize the cumulative reward.

[0096] Under the framework of Reinforcement Learning from Human Feedback (RLHF), the agent tries to maximize the reward obtained from the environment by performing actions. For an LLM, its "action" is the generated text response, and the "environment" includes the human user it interacts with and the feedback mechanism provided by the user.

[0097] Specifically, the embodiments of the present application regarding reward modeling include the following steps:

[0098] Collect feedback data: First, a series of text responses are generated from the e-commerce vertical domain LLM. These responses are presented to human evaluators, who score or classify each response according to preset criteria (such as accuracy, usefulness, etc.) to form feedback data.

[0099] Build a reward model: Based on the collected feedback data, train a reward model that can predict the reward value (or score) that any given text response should receive. The core of this step is to let the machine learn how to simulate human judgment of text quality. Among them, the reward model can be, for example, a classification model, a relevance model, a binary classification model, text classification, bert, gpt and other models.

[0100] Reinforcement learning stage: Use reinforcement learning algorithms (such as PPO, SAC, etc.), regard the e-commerce vertical domain LLM as an agent, and fine-tune it through interaction with the reward model. The agent will adjust its own strategy according to the reward signal obtained from the reward model, that is, change the way it generates text, so that the text generated in the future is more likely to obtain a higher reward.

[0101] In this way, the e-commerce vertical domain LLM fine-tuned based on reward modeling can meet expectations to a greater extent and improve the quality and accuracy of the generated content.

[0102] Taking an example, assume that the generated responses of the e-commerce vertical LLM include 10 results. These 10 results are ranked, and the top 3 are intercepted as positive examples, and the last 3 are used as negative examples. Then, the model is fine-tuned based on the positive and negative examples to make the model closer to human expectations.

[0103] Examples of positive and negative examples are as follows:

[0104] Positive example: "Wang Xiaoer Braised Chicken, Attribute 1: Wang Xiaoer - brand word, Attribute 2: Braised Chicken - category word";

[0105] Negative example: "Wang Xiaoer Braised Chicken, Attribute 1: Wang Xiaoer - category word, Attribute 2: Braised Chicken - category word";

[0106] In the above example, "Wang Xiaoer" should be a brand word, but it is wrongly regarded as a category word in the negative example, so its reward value (score) is relatively low and it is used as a negative example.

[0107] III. Construction of the External RAG Knowledge Base

[0108] Through data statistical analysis, the decision factors can be divided into several categories, including but not limited to: personal factors (needs / desires, economic status, psychological factors, cognition), product factors (quality, price, functionality, design and appearance), situational factors (shopping occasions, social culture), information communication factors (evaluations and reviews, advertising and promotions, trust and security, after-sales service), social and network factors (word-of-mouth communication, community influence, network effect), technical and interface factors (user experience, mobile compatibility, technical support).

[0109] Table 2

[0110]

[0111] Based on the statistical analysis results of the above decision factors, an external RAG knowledge base is summarized and constructed.

[0112] III. Implementation of COT-RAG

[0113] In the embodiments of this application, on the basis of the original domain knowledge base of the LLM, an external RAG knowledge base is additionally added, and the prompt is optimized by using the COT+RAG (RAT) method.

[0114] Among them, see Figure 4, showing a schematic diagram of the domain knowledge base and the external knowledge base. The domain knowledge base is a set of manually annotated data, and the external knowledge base, such as Baidu Encyclopedia, can provide data sets with the latest dynamics. Taking the search Query term "iphone16" as an example, the data given by the domain knowledge base includes: data of iphone15 related to iphone16; the data given by the external knowledge base includes: text data and structured data, specifically see Figure 4 .

[0115] COT (Chain of Thought) is a prompting technique designed to help LLMs perform step-by-step reasoning when solving complex problems. By prompting the LLMs to think step by step, COT can significantly improve the performance of the model in logical reasoning tasks. Specifically, COT is achieved in the following two ways: 1. Zero-shot prompting: By using special words or phrases (such as "Let's think step by step") in the prompt to guide the LLMs to explain their reasoning process; 2. Few-shot prompting: Show the LLMs several examples of solving similar problems, and these examples clearly explain the problem-solving steps. After observing these examples, the LLMs will try to imitate this step-by-step reasoning method.

[0116] RAG (Retrieval Augmented Generation) allows LLMs to access external information sources during the reasoning process. When the LLMs encounter problems that require specific knowledge to solve, RAG will retrieve relevant information from the external knowledge base to ensure that each step of reasoning is based on accurate knowledge. In this way, the LLMs can reduce the risk of generating hallucinatory solutions and improve the accuracy and reliability of the answers.

[0117] RAT (Retrieval Augmented Thoughts) is a combination of COT and RAG, aiming to solve the "hallucination" problem that occurs in the long-task reasoning of LLMs. RAT is achieved through the following steps:

[0118] 1. Generate the initial prompt: for CoT, encouraging step-by-step thinking.

[0119] 2. Generate the prompt2 for RAG retrieval based on the steps after splitting the initial answer: Retrieve relevant information from the external knowledge base.

[0120] 3. Integrate the results of the previous two steps into the revised prompt3: Let the LLMs give the final result.

[0121] See Figure 5, which shows a schematic diagram of the application of RAG in the vertical domain LLM training stage. The preparatory work is to construct a domain knowledge base and an external knowledge base, extract query words from the domain knowledge base, perform text segmentation on the external knowledge base, summarize the two to form a documents file library, and perform vectorization processing on the file library (for example, through the GTE vector model); obtain the user Query, perform vector retrieval in the file library according to the Query, and re-rank the retrieved results, splice the sorted retrieved results into the prompt to obtain the target prompt, and train and infer the LLM according to this target prompt, and finally realize label classification of the query.

[0122] See Figure 6 , which shows a schematic diagram of RAG optimizing the prompt in the Q&A session. First, receive the user Query and perform pre-search optimization, then perform processing such as regularization, rewriting, and query expansion on the Query. Next, perform RAG retrieval in the external RAG knowledge base, and re-rank the RAG results. Inject the RAG results into the prompt (the initial prompt generated by the prompt generator) according to the sorting results to obtain the target prompt. Finally, input the target prompt into the LLM to prompt the LLM.

[0123] A Prompt can be a text description, keywords, or other forms of instructions that guide the model to generate text that meets the requirements. It is generated by the prompt generator during the conversation with the e-commerce vertical domain LLM.

[0124] For example, the text-based Prompt can be divided into several paragraphs. For example, the prompt includes the following paragraphs:

[0125] The first paragraph: Role setting, such as "You are an expert in local life and you are very good at..."

[0126] The second paragraph: Examples of ICL, such as "Here are some examples. Example 1: query = fruit tea, user =, LBS =, and its decision factors are = xxx, xxx1. Example 2: query = drink, user =, LBS =, and its decision factors are = xxx, xxx1."

[0127] The third paragraph: The current request information, such as "Now, the information of this request is, query = milk tea, user =, LBS =..."

[0128] The fourth paragraph: Requirements for its answer and format, such as "Please answer what its decision factors are and separate them with commas."

[0129] Among them, the sample in the second paragraph and the request information in the third paragraph are both personalized texts related to the current query. The first and fourth paragraphs are generally fixed texts. The sample in the second paragraph is retrieved from the sample library by the prompt generator.

[0130] IV. Consumption Trend Recognition LLM for Sequence Modeling

[0131] In the embodiments of the present application, the consumption trend recognition LLM is a generative recommendation model, and the samples it depends on are serialized samples. Specifically, first, user population data, category data, and spatio-temporal field data are obtained. Then, serialized samples including user decision attributes, category attributes, and spatio-temporal attributes are constructed based on the user population data, category data, and spatio-temporal field data. Finally, the e-commerce vertical domain LLM is trained and inferred based on the serialized samples to obtain the consumption trend recognition LLM, and this generative model is used to perform trend analysis on consumer decision factors.

[0132] Among them, user population data: These data include users' personal information, consumption habits, preferences, etc., and are used to understand and analyze the basic characteristics of the target consumer group. Category data: These data involve the classification, characteristics, prices, etc. of goods or services, and help to understand the product supply situation in the market. Spatio-temporal field data: These data contain information in terms of time and space, such as timestamps, geographical locations, etc., and can reflect consumers' behavior patterns at different times and locations.

[0133] Perform serialization processing on the data (user population data, category data, spatio-temporal field data) to obtain sequence-style sample examples as follows:

[0134] (1): time_1 (time attribute), loc_1 (location attribute), user profile_1, query / category word_1, decision attribute of the user's click at that time_1, order amount_1;

[0135] (2) time_2 (time attribute), loc_2 (location attribute), user profile_2, query / category word_2, decision attribute of the user's click at that time_2, order amount_2;

[0136] (3) time_3 (time attribute), loc_3 (location attribute), user profile_3, query / category word_3, decision attribute of the user's click at that time_3, order amount_3;

[0137] (4) time_4 (time attribute), loc_4 (location attribute), user profile_4, query / category word_4, decision attribute of the user's click at that time_4, order amount_4;

[0138] The above four-line sequence forms a serialized sample, and the above 4 sub-elements (or more, such as 10 sub-elements) constitute a sample. In practice, serialized samples are constructed separately for commodities of different categories. For example, a milk tea serialized sample is constructed for "milk tea", and a instant noodle serialized sample is constructed for "instant noodles", and so on. The serialized samples of the above various categories of commodities are used as input signals for the e-commerce vertical domain LLM, and the vertical domain LLM is fine-tuned (for example, supervised fine-tuning SFT) to obtain a consumption trend recognition LLM.

[0139] In specific implementation, a generative recommendation model architecture (Hierarchical Sequential Transduction Units, HSTU) can be used to implement the consumption trend recognition LLM, which can effectively solve the challenges faced in large-scale recommendation systems, especially in dealing with high-dimensional, heterogeneous feature data and billions of user behavior data per day. The HSTU add-to-cart adopts a Hierarchical Sequential Transduction Units structure, which gives the model a significant speed advantage when processing long sequences. Specifically, based on the generative recommendation model architecture, a hierarchical sequential transduction unit structure is adopted to process the serialized sample to implement the consumption trend recognition LLM for sequence modeling.

[0140] For example, according to the conventional method of predicting based on historical behavior statistics results, the user behavior data (such as the number of user orders) will be statistically analyzed, and the decision factors will be sorted according to the user behavior data. This prediction method is static and one-sided, relying only on historical statistical results and not predicting future trends. However, the consumption trend recognition LLM in the embodiments of the present application does not simply sort according to historical statistical data, but combines the data serialized sample. Due to the addition of spatio-temporal data such as time position, it can predict future consumption trends, so as to realize the sorting of the decision factor set based on consumption trend prediction.

[0141] Taking an example to illustrate, assume that a user has placed few orders for coffee within half a year, only 5 times (but all concentrated in the recent 2 weeks). If according to the original method of predicting based on historical statistics, and the time window is half a year for counting the number of orders for coffee, only relying on the historical statistical result of few orders for coffee, the recognition result is that the probability of placing an order for coffee is small. However, through the consumption trend recognition LLM of the embodiments of the present application, a consumption trend prediction will be given within the time window, for example, a trend line is given. From this trend line, it can be found that although the total number of orders for coffee by this user is small, they are all concentrated in the recent two weeks. Therefore, it can be estimated that the probability of this user placing an order for coffee in the next week is still relatively high.

[0142] Thus, the consumption trend recognition LLM is obtained by further training the e-commerce vertical domain LLM based on data serialization samples, with the aim of making trend-based predictions for consumption decisions: sorting the results (decision factor sets) given by the e-commerce vertical domain LLM based on consumption trends, rather than simply sorting each decision factor based on historical behavior statistics as in the current solution.

[0143] See Figure 7 , which is a schematic diagram of the consumption decision intention analysis using two large models (e-commerce vertical domain LLM & consumption trend recognition LLM) in the embodiments of this application.

[0144] Figure 7 Describes a consumer decision intention characterization system based on LLM. The system mainly consists of the following parts: user population data, category data, spatio-temporal field data, Prompt generator, RAG recognition, e-commerce vertical domain LLM, decision factor set, consumption trend recognition LLM, statistical prediction, concentration characterization, and the final application scenario.

[0145] First, the system obtains information from three different data sources:

[0146] 1. User population data: These data include users' personal information, consumption habits, preferences, etc., and are used to understand and analyze the basic characteristics of the target consumer group.

[0147] 2. Category data: These data involve the classification, characteristics, prices, etc. of goods or services, and help to understand the product supply situation in the market.

[0148] 3. Spatio-temporal field data: These data contain information on time and space, such as timestamps, geographical locations, etc., and can reflect consumers' behavior patterns at different times and locations.

[0149] Then, the system inputs these data into the Prompt generator to generate a series of prompt words or questions to guide the e-commerce vertical domain LLM for further processing. The entity recognition module is responsible for extracting key entities from the text, such as names of people, places, brand names, etc., for subsequent analysis.

[0150] Next, the e-commerce vertical domain LLM processes the above data and the query, and outputs a decision factor set. These decision factors may include but are not limited to the following types:

[0151] 1. Consumer demand: The current demand status of consumers, such as hunger level, shopping desire, etc.

[0152] 2. Product attributes: The characteristics of the product itself, such as taste, quality, packaging, etc.

[0153] 3. Competitive environment: the situation of similar products or services provided by other businesses in the market.

[0154] 4. Time factors: Time of day, holidays and other factors may affect consumers’ purchasing behavior.

[0155] 5. Geographic location: The location of the consumer and his surroundings also influence his decision.

[0156] Then, through the consumption trend identification LLM, future consumer behavior trends can be predicted based on historical data and the current set of decision factors. This step can help merchants better plan inventory management, promotional activities, and other aspects.

[0157] Finally, the system provides support for five application scenarios: search, product selection, recommendation, advertising, and operation based on the concentration characterization results. Specifically:

[0158] 1. Search: Provide more accurate search result sorting based on the consumer’s query content and contextual information.

[0159] 2. Product selection: Help merchants choose which products should be put on sale, or adjust the combination strategy of existing products.

[0160] 3. Recommendation: Push products or services that consumers may be interested in to increase conversion rates.

[0161] 4. Advertising: Optimize advertising strategies to ensure that ads reach the target audience that is most likely to make a purchase.

[0162] 5. Operations: Provide overall operational guidance to merchants, including suggestions on pricing strategies, marketing campaign planning, etc.

[0163] In summary, the embodiment of the present application is based on the characterization of consumer decision intentions based on LLM, aiming to utilize multi-dimensional data and advanced natural language processing technology to deeply explore consumers' real needs and behavioral patterns.

[0164] In addition, it should be noted that the above two large models (e-commerce vertical domain LLM and consumer trend identification LLM) are independent and decoupled from each other. Assuming that the e-commerce vertical domain LLM and the consumer trend identification LLM are interdependent, they will both sink together without knowing it. For example, if the e-commerce vertical domain LLM chooses the wrong factors, the consumer trend identification LLM may not reject them but accept them.

[0165] In the embodiment of the present application,

[0166] E-commerce vertical domain LLM: This LLM is obtained by continuously pre-training a general LLM on the basis of combining a domain knowledge base and an external knowledge base for RAG, and is used for consumer intent recognition. The output result is a set of decision factors;

[0167] Consumer trend recognition LLM: It is obtained by training the e-commerce vertical domain LLM with time series samples, and is used for predicting and analyzing consumer intent trends, so as to sort the decision factors in the set of decision factors.

[0168] It can be seen that due to the different training samples and fine-tuning / training methods of the two, they are in an independent (decoupled) relationship, that is, the two have a mutual supervision effect, which can be called a dual-model self-supervised mechanism. For example, since the e-commerce vertical domain LLM and the consumer trend recognition LLM are independent of each other, for example, if the e-commerce vertical domain LLM selects a wrong decision factor, then the prediction score of the consumer trend recognition LLM for this wrong factor will be very low, such as 0.03 points. Therefore, this factor can be filtered out.

[0169] To sum up, there are essential differences and significant technical effects in both the training method and the recognition method between the solution of this application and the existing vertical domain LLM (obtaining the vertical domain LLM by only training the general LLM with the vertical domain data in this field).

[0170] First, not only train the e-commerce vertical domain LLM, but also further train the consumer trend recognition LLM on the basis of the e-commerce vertical domain LLM, that is, use a dual-model to recognize the consumer intent trend. The dual-model is self-supervised (the two are independent of each other and supervise each other), which can effectively improve the recognition accuracy;

[0171] Second, in the process of using the e-commerce vertical domain LLM to identify decision factors, not only based on the internal knowledge base of the model, but also adopt the RAG method to use the external RAG knowledge base for retrieval enhancement, expanding the retrieval scope to improve the retrieval accuracy;

[0172] Third, adopt a two-step recognition mode. First, use the e-commerce vertical domain LLM to make a "decision" to determine the set of decision factors for the current search term (Query), and then use the consumer trend recognition LLM to give the concentration characterization result of the set of decision factors to achieve the prediction of the consumer decision trend. Compared with the conventional method that only gives the recognition result without trend analysis, the consumer decision trend analysis of this application has guiding significance for the specific scenarios of e-commerce platforms (product selection, recommendation, advertising, operation, etc.).

[0173] In addition, the embodiments of this application also provide a device for implementing the above method. The specific implementation principle and implementation process can refer to the foregoing description, which will not be elaborated here.

[0174] An embodiment of the present application further provides a storage medium storing a computer program, wherein the computer program is configured to execute the steps in any of the above method embodiments when running.

[0175] Optionally, in this embodiment, the above storage medium may include, but is not limited to: various media capable of storing computer programs such as USB flash drives, read-only memories (ROM for short), random access memories (RAM for short), mobile hard disks, magnetic disks, or optical discs.

[0176] An embodiment of the present application further provides an electronic device, including a memory and a processor. A computer program is stored in the memory, and the processor is configured to run the computer program to execute the steps in any of the above method embodiments.

[0177] Optionally, the above electronic device may further include a transmission device and input / output devices, wherein the transmission device is connected to the above processor, and the input / output devices are connected to the above processor.

[0178] Optionally, specific examples in this embodiment may refer to the examples described in the above embodiments and optional implementation manners, and will not be elaborated herein.

[0179] The serial numbers of the above embodiments of the present application are only for description and do not represent the advantages and disadvantages of the embodiments.

[0180] In the above embodiments of the present application, the descriptions of the various embodiments have their own emphases. For parts not detailed in a certain embodiment, reference may be made to the relevant descriptions of other embodiments.

[0181] In several embodiments provided by the present application, it should be understood that the disclosed technical content can be implemented in other ways. Among them, the device embodiments described above are only illustrative. For example, the division of the units is only a logical function division. In actual implementation, there may be other division methods. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the displayed or discussed mutual coupling or direct coupling or communication connection can be through some interfaces. The indirect coupling or communication connection of units or modules can be in an electrical or other form.

[0182] The units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or may be distributed to multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0183] In addition, in each embodiment of the present application, each functional unit may be integrated into one processing unit, or each unit may exist physically alone, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of a software functional unit.

[0184] If the above-mentioned integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it may be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, may be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in each embodiment of the present application. The foregoing storage medium includes: various media such as USB flash drives, read-only memories (ROMs), random access memories (RAMs), mobile hard disks, magnetic disks, or optical discs that can store program codes.

[0185] The above are only the preferred embodiments of the present application. It should be noted that for those of ordinary skill in the art, without departing from the principle of the present application, several improvements and refinements can still be made, and these improvements and refinements should also be regarded as the protection scope of the present application.

Claims

1. A method for analyzing the consumption decision-making intention of an e-commerce platform, characterized in that Including: Based on the search term, as well as user population data, category data, and / or spatio-temporal field data, use a prompt text generator to generate an initial prompt text; Perform vector retrieval in an external RAG knowledge base based on the search term, and splice the retrieval results into the initial prompt text to obtain a target prompt text; Input the target prompt text into a pre-trained e-commerce vertical domain LLM to obtain a set of decision factors representing consumption intention decisions, where the e-commerce vertical domain LLM is obtained by continuously pre-training a general LLM based on e-commerce vertical domain data; According to a pre-trained consumption trend recognition LLM, perform trend recognition on the set of decision factors to obtain a concentration characterization result of the decision factors, where the consumption trend recognition LLM is obtained by training and reasoning on the e-commerce vertical domain LLM based on serialized samples including user decision attributes, category attributes, and / or spatio-temporal attributes constructed from user population data, category data, and / or spatio-temporal field data.

2. The method according to claim 1, wherein The performing vector retrieval in an external RAG knowledge base based on the search term and splicing the retrieval results into the initial prompt text to obtain a target prompt text includes: Receive the search term, and perform regularization, rewriting, and query expansion processing on the search term; Perform vector retrieval in an external RAG knowledge base based on the processed result of the search term, and perform chain-of-thought re-ranking on the retrieval results; Inject the retrieval results into the initial prompt text in the re-ranking order to obtain a target prompt text.

3. The method according to claim 2, wherein The method further includes: Construct an external RAG knowledge base, where one or more dimensional data of personal factors, product factors, scenario factors, information communication factors, technology and interface factors are statistically analyzed, and a set of decision factors for consumption decisions is determined based on the dimensional data to constitute the external RAG knowledge base, where the personal factors include demand / desire factors, economic status factors, psychological factors, cognitive factors; the product factors include quality factors, price factors, functional factors, design and appearance factors; the scenario factors include shopping occasion factors, social and cultural factors; the information communication factors include evaluation and review factors, advertising and promotion factors, trust and security factors, after-sales service factors; the social and network factors include word-of-mouth communication factors, community influence factors, network effect factors; the technology and interface factors include user experience factors, mobile compatibility factors, technical support factors.

4. The method according to claim 1, wherein The continuously pre-training a general LLM based on e-commerce vertical domain data includes: Obtain e-commerce vertical domain data from an in-vertical-domain material library, in-vertical-domain knowledge graph, and / or in-vertical-domain picture library of e-commerce; Perform feature extraction on the e-commerce vertical domain data to determine the data expressions of the knowledge dimension and the product dimension. The data expressions of the knowledge dimension include any one or more of attribute information, category information, and label information, and the data expressions of the product dimension include any one or more of brand, taste, merchant, and evaluation. Based on the knowledge - dimension data expression and product - dimension data expression of the e - commerce vertical domain data, training samples are determined, and the general LLM is continuously pre - trained.

5. The method according to claim 4, wherein It further includes: Performing reward modeling on the e - commerce vertical domain LLM. Among them, a series of text responses are generated from the e - commerce vertical domain LLM, scored or classified based on the text responses to form feedback data; based on the collected feedback data, a reward model is trained; using the reinforcement learning algorithm, taking the e - commerce vertical domain LLM as an agent, the e - commerce vertical domain LLM is fine - tuned through interaction with the reward model.

6. A modeling method for analyzing the consumption decision-making intention of an e-commerce platform, characterized in that, It includes: Based on the e - commerce vertical domain data, continuously pre - train the general LLM to obtain the e - commerce vertical domain LLM; Based on user population data, category data, and / or spatio - temporal field data, construct serialized samples including user decision attributes, category attributes, and / or spatio - temporal attributes, and train and infer the e - commerce vertical domain LLM based on the serialized samples to obtain the consumption trend recognition LLM; Among them, when performing consumption decision - making intention analysis, first determine the target prompt text and use the target prompt text as the input of the e - commerce vertical domain LLM to obtain a set of decision factors representing consumption intention decisions; then, according to the consumption trend recognition LLM, perform trend prediction on the set of decision factors to obtain the concentration characterization result of the decision factors; the determination process of the target prompt text includes: based on the search term, as well as user population data, category data, and / or spatio - temporal field data, use a prompt text generator to generate an initial prompt text; perform vector retrieval in the external RAG knowledge base based on the search term, and splice the retrieval result into the initial prompt text to obtain the target prompt text.

7. An apparatus for analyzing the consumption decision intention of an e-commerce platform, characterized in that, For executing the method according to any one of claims 1 - 5.

8. A modeling device for analyzing consumption decision-making intentions of an e-commerce platform, characterized in that, For executing the method according to claim 6.

9. A storage medium, characterized in that, A computer program is stored in the storage medium, where the computer program is set to execute the method according to any one of claims 1 to 6 when running.

10. An electronic device, comprising a memory and a processor, characterized in that, A computer program is stored in the memory, and the processor is set to run the computer program to execute the method according to any one of claims 1 to 6.

Images