Commodity intelligent recommendation system based on large model and intelligent recommendation system

Abstract

The embodiment of the invention provides an intelligent commodity recommendation system based on a large model, and the system is characterized in that on electronic equipment, a data collection module collects user behavior data to generate a historical behavior data sample, a natural language module processes an input historical natural language query text based on a natural language processing model, and the processed historical behavior data sample is stored in a database; extracting a user demand feature sample and an intention feature sample; on the background server, a user portrait generation module generates a user portrait according to a historical behavior data sample, a user demand feature sample and an intention feature sample; the commodity recommendation module matches the user portrait with the commodity feature library to generate a recommendation result, and the result generation module generates multi-screen push content according to the recommendation result and the display size of the reference electronic equipment, pushes the multi-screen push content to the user one by one, collects feedback data of the user on the current screen push content, and sends the feedback data to the user. And the commodity recommendation module optimizes the multi-screen push content and pushes the multi-screen push content again.

Description

Technical Field

[0001] The embodiments of the present invention relate to the field of data processing, and in particular to a commodity intelligent recommendation system and an intelligent recommendation system based on a large model. Background Art

[0002] In today's digital business environment, product recommendation systems are crucial for e-commerce platforms and various sales companies, directly impacting user shopping experiences and company sales performance. However, existing product recommendation systems face numerous challenges that need to be addressed.

[0003] Traditional product recommendation systems mostly rely solely on simple data analysis, such as users' past purchase history and browsing behavior. This approach captures a single dimension of data and fails to deeply understand users' true needs and underlying intentions. For example, simply recommending a brand of sneakers or similar styles based solely on a user's past purchase of that brand ignores other related needs (such as sports backpacks and water bottles) arising from specific scenarios (e.g., outdoor activities), making it difficult to meet users' diverse and personalized shopping needs.

[0004] Furthermore, existing recommendation systems have limited capabilities for processing natural language user queries. When users enter complex or ambiguous queries in natural language, the system cannot accurately extract the user's needs and intent. For example, if a user enters "I want a lightweight and fashionable summer outfit," the system will only recognize simple keywords like "apparel" and "summer," but will not accurately grasp key demand characteristics like "lightweight" and "fashionable," resulting in recommended products that are far from the user's expectations.

[0005] Furthermore, the recommendation results generated by existing recommendation systems lack adaptability to the display sizes of different devices. In the multi-screen era, users browse products through a variety of devices, including mobile phones, tablets, and computers, and the screen sizes of different devices vary greatly. Existing recommendation systems fail to generate appropriate push content based on the device display size, resulting in incomplete information display on small-screen devices and inconsistent layouts on large-screen devices, seriously affecting the user viewing experience. At the same time, existing systems do not adequately collect and utilize user feedback data, making it impossible to optimize recommendations in a timely manner based on user feedback on push content, making it difficult to continuously improve recommendation effectiveness.

[0006] In summary, the shortcomings of current product recommendation systems in data collection and analysis, natural language processing, multi-screen adaptation, and user feedback optimization have greatly limited the accuracy and effectiveness of their recommendations, making it impossible to provide users with high-quality, personalized product recommendation services. The large-scale model-based intelligent product recommendation system provided in the embodiments of this application is expected to effectively solve the problems existing in the above-mentioned existing technologies. Summary of the Invention

[0007] In view of this, embodiments of the present invention provide a large-model-based intelligent product recommendation system and an intelligent recommendation system to at least partially solve the above problems.

[0008] According to a first aspect of an embodiment of the present invention, a large-model-based intelligent product recommendation system is provided, which includes: an electronic device and a backend server, wherein the electronic device is configured with a data collection module and a natural language module, the data collection module is used to collect user behavior data to generate historical behavior data samples, and the natural language module is used to process the input historical natural language query text based on a natural language processing model to extract user demand feature samples and intention feature samples; the backend server is configured with a user portrait generation module, a product recommendation module, and the result generation module, the user portrait generation module generates a user portrait based on historical behavior data samples, user demand feature samples and intention feature samples; the product recommendation module is used to match the user portrait with a product feature library to generate a recommendation result, and the result generation module is used to generate multi-screen push content based on the recommendation result and with reference to the display size of the electronic device and push it to the user one by one, and at the same time, collect user feedback data on the current one-screen push content, so that the product recommendation module optimizes the multi-screen push content accordingly and pushes it again.

[0009] According to a second aspect of an embodiment of the present invention, a large-scale model-based intelligent recommendation system is provided, comprising: an electronic device and a backend server. The electronic device is configured with a data collection module and a natural language module, wherein the data collection module is configured to collect user behavior data to generate historical behavior data samples; the natural language module is configured to process input historical natural language query text based on a natural language processing model to extract user demand feature samples and intent feature samples;

[0010] The backend server is equipped with a user profile generation module, a recommendation object matching module, and a result generation module. The user profile generation module generates a user profile based on historical behavior data samples, user demand feature samples, and intention feature samples; the recommendation object matching module is used to match the user profile with the recommendation object feature library to generate recommendation results; and the result generation module is used to generate push content based on the recommendation results and push it to the user.

[0011] In the solution of the embodiment of the present invention, the intelligent product recommendation system based on the large model has the following technical advantages:

[0012] 1. The system of the present application collects user behavior data through the data collection module on the electronic device to generate historical behavior data samples. At the same time, the natural language module processes the historical natural language query text based on the natural language processing model to extract user demand feature samples and intention feature samples. Compared with the traditional method of relying solely on simple data such as historical purchase records and browsing behavior, this broadens the data dimension. For example, user demand features such as "suitable for summer wear, lightweight and fashionable" extracted by the natural language module can enable the system to understand the user's more detailed and diverse needs, rather than just being limited to the categories of goods purchased in the past. These multi-dimensional data provide rich information for the subsequent generation of more comprehensive and accurate user portraits, which helps to deeply understand the user's real needs and potential intentions, thereby meeting the user's diverse and personalized shopping needs.

[0013] 2. The natural language module processes historical natural language query text input based on a natural language processing model, extracting samples of user demand and intent characteristics. Unlike traditional systems that can only recognize simple keywords, this module can deeply analyze user input text and accurately grasp key demand characteristics such as "lightweight" and "fashionable." This makes the system more accurate when processing complex or ambiguous natural language queries, and can convert users' natural language expressions into effective information that can be used for recommendations. This greatly improves the system's accuracy in understanding user needs and can recommend products that better meet user expectations.

[0014] 3. The result generation module generates multi-screen push content based on the display size of the device. In the multi-screen era, screen sizes vary significantly across devices. This feature ensures that recommended content is appropriately displayed on a variety of devices, including mobile phones, tablets, and computers. For example, on small-screen devices, a reasonable layout can be used to avoid incomplete information display, while on large-screen devices, the layout can be optimized to ensure consistency. This significantly improves the user's viewing experience across different devices. Regardless of the device used to browse products, users can enjoy a good visual and interactive experience, which increases their willingness to interact with the system.

[0015] 4. When pushing content, the result generation module collects user feedback data on the content pushed on the current screen. The product recommendation module optimizes the multi-screen push content based on this feedback data and pushes it again. Unlike traditional systems that do not collect and utilize user feedback data, this system can adjust the recommendation strategy in a timely manner based on user feedback. For example, if a user feedbacks that the recommended product on a certain screen does not meet their needs, the system can adjust the type and style of subsequent recommended products based on this feedback, continuously optimizing the recommendation effect, making the recommendation results increasingly in line with the user's actual needs, thereby providing users with higher-quality, personalized product recommendation services and improving user satisfaction and loyalty to the recommendation system. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments recorded in the embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings.

[0017] Figure 1 This invention provides an intelligent commodity recommendation system based on a large model. DETAILED DESCRIPTION

[0018] Figure 1 The embodiment of this application provides a commodity intelligent recommendation system based on a large model. Figure 1 As shown, the intelligent product recommendation system based on the large model includes: an electronic device and a background server, the electronic device is configured with a data collection module and a natural language module, the data collection module is used to collect user behavior data to generate historical behavior data samples, the natural language module is used to process the input historical natural language query text based on the natural language processing model to extract user demand feature samples and intention feature samples; the background server is configured with a user portrait generation module, a product recommendation module, and the result generation module, the user portrait generation module generates a user portrait based on historical behavior data samples, user demand feature samples and intention feature samples; the product recommendation module is used to match the user portrait with the product feature library to generate a recommendation result, the result generation module is used to generate multi-screen push content based on the recommendation result and refer to the display size of the electronic device and push it to the user one by one, and at the same time, collect user feedback data on the current one-screen push content, so that the product recommendation module optimizes the multi-screen push content accordingly and pushes it again.

[0019] Optionally, the data collection module is also used to collect historical recommendation content data, the product recommendation module is also used to adjust the recommendation results based on the historical recommendation content data, and the result generation module is used to generate multi-screen push content based on the adjusted recommendation results and with reference to the display size of the electronic device.

[0020] Optionally, when collecting user behavior data and / or historical recommended content data, the data collection module detects the user's behavior records on web pages and APPs based on the set crawler component to collect user behavior data and / or historical recommended content data and store them in the basic database according to the set storage mechanism.

[0021] Optionally, the data collection module is deployed with a data collection framework to trigger the collection of user behavior data and / or historical recommended content data based on the data collection framework.

[0022] Preferably, in a specific application scenario, the above solution is described as follows in a preferred or alternative manner.

[0023] 1. Data Collection Module

[0024] The data collection module is responsible for collecting user behavior data (UBD) and historical recommended content data (HRCD). This module performs data collection based on crawler components and a data collection framework.

[0025] Let P UBD (t) represents the probability of collecting user behavior data at time t, which is affected by many factors, such as the user's activity level A on the web page or APP user (t), the efficiency coefficient α of the crawler component, the trigger frequency f of the data collection framework trigger (t), etc. This application defines the following formula to describe:

[0026]

[0027] in, Indicates different types of user behaviors (such as browsing, clicking, purchasing, etc., respectively with S i (t) represents, β i is the weight coefficient of the corresponding behavior type) on the comprehensive impact of the collection probability. For example, the purchase behavior S purchase The weight β of (t) purchase Relatively high, because it can better reflect the real needs of users.

[0028] Similarly, for the collection probability P of historical recommended content data HRCD HRCD (t), this application has:

[0029] Here, A rec (t) represents the activity level of historical recommended content, γ is the collection efficiency coefficient related to historical recommended content, R j (t) represents different types of historical recommendation content (such as recommended product categories, recommended copywriting styles, etc.), δ j is its corresponding weight coefficient.

[0030] The collected data is stored in the basic database according to the set storage mechanism. Let the relationship between storage capacity C and time t be:

[0031]

[0032] Among them, C0 is the initial capacity of the basic database, V UBD and V HRCD They represent the storage space required for unit user behavior data and unit historical recommended content data respectively.

[0033] 2. Product recommendation module

[0034] The product recommendation module adjusts the recommendation results based on historical recommendation content data. original is the initial recommendation result, R adjusted This is the adjusted recommendation result. This application is adjusted using the following formula:

[0035]

[0036] Among them, λ is the adjustment coefficient, which is used to control the influence of historical recommendation content data on the recommendation results. k represents the k-th category of historical recommended content data, is the average value of all historical recommended content data, ω k is the weight of the k-th category of historical recommendation content data. For example, if a certain category of historical recommendation content (such as a recommended product combination with a high conversion rate) has a high correlation with the current recommendation result, its corresponding ω k The value is larger.

[0037] 3. Result generation module

[0038] The result generation module is based on the adjusted recommendation result R adjusted , refer to the electronic device display size D to generate multi-screen push content. Assume that the optimization degree O of multi-screen push content is:

[0039]

[0040] Among them, S represents the number of different screen types (such as mobile phone screens, tablet screens, computer screens, etc.), D s is the display size of the sth screen, F(R adjusted ,D s ) is the actual push content feature generated based on the adjusted recommendation results and the sth screen size, I s It is the ideal push content feature of the sth screen, M s is a normalization parameter, μ s is the importance weight of the sth screen in multi-screen push notifications. For example, mobile screens have a relatively high value due to their more frequent use. Using this formula, the result generation module can optimize push content based on different screen sizes to provide a better user experience.

[0041] To this end, the technical benefits of the above preferred solution are described as follows:

[0042] 1. Technical benefits of the data collection module formula

[0043] Accurately characterize the collection probability: By introducing and Such a formula can accurately take into account many factors that affect the probability of collecting user behavior data and historical recommendation content data. For example, the user's activity level A on the web page or APP user Including (t) in the calculation means that the system can dynamically perceive the real-time participation of users. The higher the activity, the greater the probability of collecting valid data, which helps to capture the changes in user behavior at different times. i (t) and its weight β i The setting enables the system to collect data with emphasis according to the importance of the behavior, such as purchase behavior S purchase (t) The higher weight ensures the priority collection of key user behaviors, thus providing a more valuable data basis for subsequent recommendations.

[0044] Dynamically managing storage capacity: Formula This formula clearly describes the dynamic changes in the basic database storage capacity over time. It considers the collection probability of different types of data (user behavior data and historical recommendation content data) and their unit data storage space requirements, enabling the system to plan and manage storage resources in advance. For example, if user activity is predicted to increase over a certain period, leading to an increase in data collection volume, the system can prepare sufficient storage space in advance based on this formula, avoiding the loss of important data due to insufficient storage and ensuring the continuity and integrity of data collection.

[0045] 2. Technical Benefits of the Product Recommendation Module Formula

[0046] Optimizing recommendations based on historical data: Formula This demonstrates the advantage of using historical recommendation content data to intelligently adjust the initial recommendation results. By analyzing historical recommendation content data HRCD k and the average The difference, combined with the weight ω k and adjustment coefficient λ, the system can mine the potential patterns and user preferences in historical recommendations. For example, if a certain type of historical recommendation content (such as recommendations for a specific product combination) has always had a high conversion rate, that is, HRCD k Large and weight ω k If the value of this type of recommendation is high, the system will increase the proportion of this type of recommendation content accordingly in the current recommendation, making the recommendation results more in line with the user's actual needs, improving the accuracy and effectiveness of the recommendation, and thus increasing the user's acceptance of the recommended products and willingness to purchase.

[0047] 3. Technical benefits of the result generation module formula

[0048] Multi-screen adaptation and optimization: formula Fully consider the display size of different electronic devices D sThe impact on push content is achieved by optimizing the push content on multiple screens. By comparing the actual push content features F(R) generated based on the adjusted recommendation results and different screen sizes, adjusted ,D s ) and the ideal push content feature of the screen I s The difference, combined with the importance weight μ of each screen s and normalization parameter M s , the system can adaptively optimize pushed content for different screen types (such as mobile phones, tablets, and computers). For example, mobile screens are smaller and require more concise and focused information display. The system will adjust the layout and presentation of pushed content based on this formula to ensure that users can quickly access key information on mobile screens, improve the user browsing experience on different devices, and enhance the user's interactivity and satisfaction with the recommendation system.

[0049] Optionally, the backend server is further configured with a model training module for performing forward propagation and backpropagation calculations on model parameters of the natural language processing model to be trained based on the natural language query text sample until a set model training end condition is reached, so as to obtain the trained natural language processing model;

[0050] When the natural language processing model processes the input historical natural language query text, it extracts user demand feature samples and intention feature samples therefrom through logical reasoning.

[0051] Preferably, the above solution is described below in a preferred or alternative manner.

[0052] The new natural language processing model proposed in this application consists of the following main structural layers: Input Embedding Layer, Hierarchical Attention Fusion Layer, Semantic Reasoning Layer, and Feature Output Layer. Each structural layer works together to accurately extract user demand feature samples and intent feature samples from historical natural language query text.

[0053] (1) Input Embedding Layer

[0054] This layer maps each word or subword in the input historical natural language query text to a low-dimensional vector space in preparation for subsequent processing.

[0055] Assume that the input natural language query text is T = [w1,w2,...,w n], where w i This application defines the embedding function E:w i →e i , each w i Mapped to a d-dimensional embedding vector e i .

[0056]

[0057] Among them, v j is the basis vector in the pre-trained word vector space, and m is the number of basis vectors. ij Is a weight coefficient, indicating the j-th basis vector w i The contribution of the embedding vector, which is determined by a context-based function f context (w i ,T) is determined, the function takes w into account i Contextual information such as position in text T, surrounding words, etc. i is a bias vector used to adjust the overall offset of the embedding vector, which is related to w i The inherent attributes such as part of speech and word frequency are related to the function g property (w i ). In this way, the embedding vector not only contains the semantic information of the word itself, but also incorporates information about the context and inherent properties of the vocabulary, making it more suitable for subsequent processing.

[0058] (2) Hierarchical Attention Fusion Layer

[0059] This layer aims to capture semantic information at different levels in the text and fuse this information through the attention mechanism.

[0060] 1. Word-level attention

[0061] First, the attention weight is calculated at the word level. For the embedding vector sequence E=[e1,e2,...,e n ], this application defines word-level attention scores Indicates the degree of attention of the i-th word to the j-th word.

[0062]

[0063] in, is the weight matrix, b w is the bias vector, q w is the query vector used to calculate the attention score. These parameters are learned through model training. Word-level attention weighted vector for:

[0064]

[0065] 2. Sentence-level attention

[0066] The word-level attention weighted vectors are combined into sentence representations, and the attention mechanism is applied again at the sentence level. Assume that the text is divided into s sentences, and each sentence S l Composed of a series of word-level attention weighted vectors Composition, i l Represents sentence S l The index of the word in the sentence. The sentence representation vector s l for:

[0067]

[0068] Calculating sentence-level attention scores Indicates the degree of attention of the lth sentence to the mth sentence: in, is the weight matrix, b s is the bias vector, q s Is the query vector used to calculate the sentence-level attention score. Sentence-level attention weight vector h s for:

[0069] Through the hierarchical attention mechanism, the model can capture important information at different levels in the text, from words to sentences, and effectively integrate them to provide richer feature representations for subsequent semantic reasoning.

[0070] (3) Semantic Reasoning Layer

[0071] This layer is based on the feature representation output by the hierarchical attention fusion layer. It mines the semantic relationships in the text through complex logical reasoning to extract user demand features and intention features.

[0072] This application introduces a reasoning structure based on knowledge graph. Assume that the knowledge graph G = (V, E), where V is a set of nodes representing various concepts and entities, and E is a set of edges representing the relationships between concepts and entities.

[0073] For the vector h output by the hierarchical attention fusion layer s , this application projects it into the vector space of the knowledge graph to obtain the projection vector p.

[0074] p=W p h s +b p , where W p is the projection weight matrix, b pis the bias vector.

[0075] Then, we perform a random walk on the knowledge graph to infer the semantic relationship. Let the current node be v i , from node v i Transfer to node v j The probability P(v j |v i ) is defined as:

[0076]

[0077] Among them, u j and u k They are node v j and v k The vector representation of N(v i ) is the node v i The set of neighbor nodes.

[0078] After t steps of random walk, we get a node sequence This application uses an aggregation function A to integrate the information of these nodes to obtain a semantic reasoning result vector r.

[0079]

[0080] Among them, σ is the activation function, θ k is the weight coefficient, which indicates the importance of the k-th random walk reaching the node, which is a function based on the node type and the walk path. Sure.

[0081] (4) Feature Output Layer (FeatureOutputLayer)

[0082] This layer outputs user demand feature samples and intention feature samples based on the results of the semantic reasoning layer.

[0083] Assuming the semantic reasoning result vector is r, this application obtains the user demand feature sample vector f through two linear transformations respectively d And the intention feature sample vector f i .

[0084] f d =W d r+b d

[0085] f i =W i r+b i

[0086] Among them, W d ,W iis the weight matrix, b d ,b i is the bias vector. After further post-processing (such as normalization and clustering), these two vectors can be used to obtain the final user demand feature samples and intention feature samples.

[0087] To this end, the advantages of the above preferred or alternative technical solutions are as follows:

[0088] In the application scenario of a large-scale model-based intelligent product recommendation system, the formulas at each structural layer of the natural language processing model described above bring many technical benefits:

[0089] Technical Benefits of Input Embedding Formula

[0090] Context and Lexical Attribute Fusion: Formula By using a context-based function f context (w i ,T) determine α ij , and functions g related to the intrinsic properties of vocabulary property (w i )Calculate β i , so that each word's embedding vector not only contains the semantic information of the general word vector space, but also incorporates the context and the characteristics of the word itself. In the product recommendation scenario, users' natural language queries are very flexible. For example, "I bought the same brand of sports backpack last time. Recommend some similar ones." This formula allows the model to fully capture the contextual information of "the brand I bought last time" and the lexical attributes of "sports backpack," thereby more accurately understanding user needs and laying a solid foundation for subsequent recommendations.

[0091] Flexible Adaptation to Different Query Formulations: By taking into account context and lexical attributes, the model is more adaptable to natural language queries expressed in various ways. Different users may use different vocabularies or expressions to describe their needs for the same product. This formula enables the model to accurately map these diverse expressions to appropriate embedding vectors, preventing misunderstanding of user intent due to differences in expression. This improves the model's tolerance and accuracy for a wide range of natural language inputs.

[0092] Technical benefits of the hierarchical attention fusion layer formulation

[0093] Word-level attention focuses on key information: Word-level attention score formula This allows the model to focus on the relationships between words in a text at the word level, highlighting key words. For example, in a natural language query for product recommendations, such as "I want a cost-effective and lightweight laptop," the model can use word-level attention to focus on key information such as "cost-effective," "lightweight," and "laptop," rather than treating each word equally, thereby accurately capturing the user's core product needs.

[0094] Sentence-level attention grasps the overall semantics: The sentence-level attention mechanism calculates the attention scores between sentences It can integrate information across different sentences and grasp the overall semantics of the text. Complex user queries often contain multiple sentences describing needs and intents, such as "I travel frequently. I want a lightweight, portable laptop with long battery life and, ideally, good heat dissipation." Sentence-level attention can integrate this information across these sentences to understand the user's need for various laptop features stemming from frequent travel. This avoids neglecting the overall semantics by focusing on a single sentence, and allows for comprehensive and accurate extraction of user needs and intents.

[0095] Technical Benefits of Semantic Reasoning Layer Formula

[0096] Knowledge graph enhances semantic understanding: Projecting the output of the hierarchical attention fusion layer into the knowledge graph vector space to obtain the formula p = W p h s +b p , and the semantic reasoning process based on random walks of the knowledge graph, enable the model to use the rich semantic relationships of the knowledge graph for reasoning. In the product recommendation scenario, the knowledge graph contains various product concepts, attributes, and the relationships between them. For example, when a user queries "shoes suitable for running", the model can infer the relationship between "running shoes" and related concepts such as "sports equipment", "comfortable insoles", and "good grip" through the knowledge graph, thereby more comprehensively understanding user needs and recommending not only ordinary running shoes, but also accessories related to running shoes, improving the richness and accuracy of recommendations.

[0097] Random walk to mine potential intentions: Based on the semantic reasoning process of random walk, the transition probability P(v j |v i ) and aggregation function A can mine the underlying semantic relationships and user intent within the text. In product recommendations, user intent isn't directly expressed but rather hidden within semantic relationships. For example, if a user says, "I like that kind of minimalist furniture," the model uses random walks on the knowledge graph to explore potential connections between "minimalist furniture" and related concepts like "modern design" and "eco-friendly materials." This reveals that the user is also interested in furniture with these characteristics, thereby providing product recommendations that better meet the user's potential needs.

[0098] Technical Benefits of the Feature Output Layer Formula

[0099] Accurately output demand and intention features: through two linear transformation formulas f d =W d r+b d and f i =W i r+bi Get the user demand feature sample vector f respectively d And the intention feature sample vector f i , accurately mapping semantic reasoning results to the user demand and intent feature space. In product recommendation systems, this allows the model to clearly distinguish and extract users' specific demand characteristics for products (such as product attributes and functions) and purchase intention characteristics (such as urgency and budget range), providing a clear basis for subsequent accurate product recommendations based on these characteristics, greatly improving the relevance and effectiveness of recommendations.

[0100] Optionally, the natural language processing model is a natural language processing component. When the natural language processing model processes the input historical natural language query text, it extracts user demand feature samples and intention feature samples from it through text semantic analysis. The text semantic analysis includes at least one of word segmentation, part-of-speech tagging, named entity recognition, and syntactic analysis.

[0101] Preferably, the above technical solution is described as follows in an alternative or preferred manner.

[0102] 1. Tokenization

[0103] Word segmentation is the process of dividing the input natural language query text T into words or subwords (called tokens). Assume that the input text T = [c1, c2, ..., c N ], where c i is the i-th character in the text.

[0104] This application defines a probability-based word segmentation function Tokenize(T), which outputs a tokens sequence w = [w1, w2, ..., w M ]. For each token boundary k (i.e. from c k to c k+l This application calculates the probability P(w m ):

[0105]

[0106] in:

[0107] σ is the sigmoid function, which is used to map the output value to the interval [0, 1] to represent the probability.

[0108] α i Is with the character c iThe relevant weight parameter reflects the contribution of each character to the composition of the token. In the product recommendation scenario, some characters have special meanings in specific product vocabulary, such as "+" has a specific meaning in the model of electronic products. In this case, α i It will be given corresponding weight.

[0109] f char (c i ) is a character feature function that extracts character c i Features such as character type (letters, numbers, punctuation, etc.), whether it is the starting character of a specific product vocabulary, etc. In the product query "I want an iPhone 14 Pro + phone case", the function can recognize characters such as "i" and "P" as features of the starting characters of the brand name.

[0110] β is a contextual weight parameter that controls the degree to which contextual information influences token classification. Contextual information is crucial for product recommendations. For example, considering the word "buy" in "I want to buy a phone" and "I bought a phone case" can more accurately determine whether it is an independent token.

[0111] f context (c ks ,...,c k+t ) is the context feature function, which comprehensively considers the ks to c k+t contextual information, such as the local phrase structure composed of several characters before and after, and whether it conforms to common product description patterns.

[0112] If P(w m ) is greater than a certain threshold τ, then c k to c k+l It is judged to be a token m .

[0113] 2. Part of Speech Tagging

[0114] Part-of-speech tagging is to mark the part of speech for each token. For the token sequence w after word segmentation, this application defines a part-of-speech tagging function POSTag(w), which outputs a part-of-speech tag sequence t = [t1, t2, ..., t M ], where t m It is w m The part-of-speech tag.

[0115] This application uses a neural network-based method to calculate the probability of part-of-speech tags. Let h m It is w m The word vector representation of (which can be obtained through the embedding layer mentioned above), and its context word vector hmu ,...,h m+v .

[0116]

[0117] in:

[0118] W i This is a weight matrix used to transform word vectors at different positions, capturing the influence of contextual information on part-of-speech judgment. In product recommendation scenarios, the word "buy" has different parts of speech in different contexts (such as "I want to buy a phone" and "The phone I bought is good"). These weight matrices can be used to learn this contextual dependency.

[0119] b is the bias vector.

[0120] q is a learnable query vector used to compute the probability distribution.

[0121] is the set of all part-of-speech tags.

[0122] Select the part-of-speech tag with the highest probability as w m Part-of-speech tagging m In product recommendation applications, accurate part-of-speech tagging helps understand the grammatical structure of user queries. For example, it can identify parts of speech such as nouns (such as "mobile phone" and "computer") and verbs (such as "buy" and "recommend"), providing a basis for subsequent analysis of user needs and intentions.

[0123] 3. Named Entity Recognition (NER)

[0124] Named entity recognition aims to identify specific entities in text, such as product names, brands, etc. For a token sequence w and a part-of-speech tag sequence t, this application defines a named entity recognition function NER(w,t), which outputs a named entity sequence e = [e1, e2, ..., e L ], where e l is the recognized named entity.

[0125] This paper constructs a conditional random field (CRF) model for named entity recognition. Let x = (w, t) be the input feature sequence and y be the named entity label sequence.

[0126]

[0127] in:

[0128] Z(x) is a normalization factor that ensures that the probability distribution sums to 1.

[0129] λ kis the characteristic function f k The relevant weight parameters are learned through training. In the product recommendation scenario, different types of named entities (such as brand names and product categories) have different feature function weights to highlight their identification features.

[0130] f k (y m ,y m1 ,x,m) is the feature function, which comprehensively considers the named entity label y at the current position m m , the label y of the previous position m1 and input feature sequence x. For example, the feature function can capture the feature that a token is the beginning of a brand name, or that a token, together with the preceding and following tokens, forms the name of a specific product category. In the context of "I want an Apple phone," the feature function can identify "Apple" as a brand entity.

[0131] By maximizing P(y|x), we find the optimal named entity tag sequence to determine the named entity e. In product recommendation applications, accurately identifying named entities is crucial for understanding the product objects that users are interested in. It can directly extract key information such as the brand and specific product name of the product that the user wants.

[0132] 4. Syntactic Analysis

[0133] Syntactic analysis is used to analyze the grammatical structure of a sentence and determine the dependency relationship between words. For a token sequence w, a part-of-speech tag sequence t, and a named entity sequence e, this application defines a syntactic analysis function SyntacticAnalyze(w, t, e), which outputs a dependency tree.

[0134] This application uses a neural network-based method, such as a Graph Neural Network (GNN), to construct a dependency tree. Suppose this application has a graph G = (V, E), where nodes V correspond to tokens and edges E represent dependencies.

[0135] For each node v m (corresponding to tokenw m ), its hidden state Update at layer t by:

[0136] Where W1 and W2 are weight matrices, used to aggregate neighbor node information and the node's own information, respectively. In product recommendation scenarios, different dependencies (such as "modification" and "subject-predicate") influence the update of the node's hidden state through these weight matrices. b is the bias vector.

[0137] After multiple layers of propagation, this application uses a classifier to predict the type of dependency relationship between each node.

[0138]

[0139] in:

[0140] r mn Represents node v m and v n The type of dependency relationship between them.

[0141] and It is the hidden state of the node after propagation through T layers.

[0142] W r is a weight matrix that maps the hidden states of two nodes to the dependency probability space.

[0143] b r is the bias vector.

[0144] q is the query vector.

[0145] Is the set of all dependency types.

[0146] By determining the dependency type, this application constructs a dependency tree In product recommendation applications, the dependency tree obtained from syntactic analysis helps understand the grammatical relationship between the various parts of a user query. For example, in the query "I want [product: mobile phone] from [brand: Apple]," the dependency tree can clarify the modification relationship of "Apple" to "mobile phone," further accurately grasping user needs.

[0147] 5. Extraction of user demand feature samples and intention feature samples

[0148] Based on the results of the above text semantic analysis (word segmentation, part-of-speech tagging, named entity recognition and syntactic analysis), this application extracts user demand feature samples and intention feature samples.

[0149] Let F d is the user demand feature sample vector, F i is the intent feature sample vector.

[0150]

[0151] in:

[0152] γ l is a named entity e lThe weight of user demand feature samples reflects the importance of different named entities in user needs. In product recommendations, the named entities corresponding to product names have higher weights.

[0153] δ m is the word vector The weight of the user demand feature sample (the word vector updated by the syntactic analysis layer) takes into account the impact of the word itself and its role in the sentence grammatical structure on user demand.

[0154] ∈ m is the word vector The weight of the intent feature samples is used to capture the information of the words in expressing the user's intent.

[0155] ω mn is the dependency relationship mn The weight of the intent feature samples can be used to infer the user's intention through the dependency relationship. For example, the dependency relationship between "want" and "product" helps to determine the user's purchase intention.

[0156] ε is the dependency tree The set of edges in .

[0157] To this end, in the specific application scenario of a product recommendation system based on natural language processing, the above technical solution is compared with traditional natural language processing technology. The significant technical benefits brought by its innovation are as follows:

[0158] 1. Participle

[0159] Traditional word segmentation methods are often based on fixed rules or simple statistical models, such as dictionary matching or ngram models. These methods struggle to cope with the complex and ever-changing natural language expressions used in product recommendation scenarios. For example, they are prone to word segmentation errors when encountering new product names, internet buzzwords, or domain-specific abbreviations.

[0160] In this application, the probability-based word segmentation formula proposed in this application Innovatively combines the character feature function f char (c i ) and context feature function f context (c ks ,...,c k+t In product recommendation scenarios, this enables the model to dynamically adapt to new product terms and complex descriptions. For example, in the sentence "I want a foldable phone with 5G support," traditional methods inaccurately segment emerging terms like "5G" and "foldable screen." However, the new technology, by considering character and contextual features, can accurately identify these terms as independent tokens, providing an accurate foundation for subsequent analysis and improving the ability to handle complex product descriptions.

[0161] 2. Part-of-speech tagging

[0162] Traditional part-of-speech tagging techniques rely heavily on predefined rules and simple probabilistic models, and therefore perform poorly when dealing with context-sensitive part-of-speech tagging. In natural language text used for product recommendations, the same word may have different parts of speech in different contexts, making it difficult for traditional methods to accurately capture this variability.

[0163] In this application, the formula for calculating the probability of part-of-speech tags based on neural networks is By considering the context information of word vectors (h mu ,...,h m+v ), can better handle context-sensitive part-of-speech tagging. For example, in the sentence "I want to learn about the specifications of this phone before deciding whether to buy it," the part-of-speech of the word "buy" needs to be determined based on the context. This new technology can accurately tag its part-of-speech using the surrounding word vector information, helping to more accurately understand the user's sentence structure and intent. Compared with traditional methods, part-of-speech tagging is more accurate in complex contexts.

[0164] 3. Named Entity Recognition

[0165] Traditional named entity recognition technologies, such as rule-based and dictionary-based methods, have limited capabilities for identifying newly emerging entities such as product brands and models, and struggle to adapt to texts from diverse domains and genres. In product recommendation scenarios, new brands and unique product names are constantly emerging, and traditional methods are unable to keep up with the times.

[0166] In this application, the named entity recognition formula based on the conditional random field (CRF) model By comprehensively considering multiple characteristic functions f k (y m ,y m1 ,x,m) and the learned weight parameter λ k , enabling more flexible recognition of various named entities. In product recommendations, whether it's a new niche brand or a complex product model, the feature function can capture its characteristics and accurately identify it. For example, in the query "I want a Dell XPS 13 Plus laptop," the new technology can accurately identify "Dell" as the brand and "XPS 13 Plus" as the specific model, demonstrating superior recognition of new and complex entities compared to traditional methods.

[0167] 4. Syntactic Analysis

[0168] Traditional syntactic analysis methods, typically based on fixed grammatical rules and shallow statistical models, are ineffective when dealing with complex, long sentences and non-standard grammatical structures. In product recommendation scenarios, users' natural language queries contain a variety of complex sentence structures and omitted structures, making it difficult for traditional methods to accurately parse them.

[0169] In this application, the formula for syntactic analysis using graph neural network (GNN) is used to update the node hidden state. and dependency prediction formula It can better handle complex grammatical structures. For example, for long sentences like "I previously bought a mobile phone with a great camera, and now I want to buy a similar one but more affordable," the new technology can use GNN to model the relationships between nodes, accurately analyze the dependency structure of the sentence, and understand the logical relationship between each part. Compared with traditional methods, it has greater advantages in handling complex long sentences and non-standard grammatical structures.

[0170] 5. Extraction of user demand feature samples and intention feature samples

[0171] Traditional methods for extracting user needs and intentions often rely solely on keyword matching or shallow semantic analysis, making it difficult to fully and accurately capture users' true needs and intentions. In product recommendation scenarios, user needs and intentions are embedded in complex semantic relationships, making traditional methods unable to deeply explore them.

[0172] In this application, the formula for extracting feature samples is obtained by synthesizing the results of each step of text semantic analysis. and This approach fully considers multiple aspects of information, including named entities, word vectors, and dependency relationships. This enables the model to more comprehensively and accurately extract user demand characteristics (such as product brand and features) and intent characteristics (such as purchase intention and price preference) for product recommendations. For example, by identifying product brands through named entities and inferring user price requirements based on dependency relationships derived from syntactic analysis, this approach can more deeply explore users' potential needs and intentions than traditional methods, providing a more reliable basis for accurate product recommendations.

[0173] Optionally, the user portrait generation module performs the following steps when generating the user portrait:

[0174] Cleaning and converting the historical behavior data samples, user demand feature samples, and intention feature samples, removing invalid samples therefrom, and obtaining valid historical behavior data samples, valid user demand feature samples, and valid intention feature samples;

[0175] Extracting features from the valid historical behavior data samples, the valid user demand feature samples, and the valid intention feature samples to obtain user behavior features, user demand features, and user intention features;

[0176] The user behavior characteristics, user demand characteristics, and user intention characteristics are quantified and classified to generate a user profile.

[0177] Preferably, the above solution is described below in an alternative or preferred manner.

[0178] 1. Data cleaning and conversion

[0179] Assume that the historical behavior data sample set is The user demand feature sample set is The intention feature sample set is

[0180] Define the data cleaning function Clean(X). For any sample set X, remove invalid samples in the following way.

[0181] For each sample x j ∈X, calculate its effectiveness score S(x j ):

[0182] in:

[0183] α k is the characteristic function f k (x j ) related weight parameters. In the product recommendation scenario, for historical behavior data samples, if the integrity of purchase behavior is crucial to judging the validity of the sample, then the f related to the integrity of purchase behavior k (x j ) corresponding to α k The weight will be higher; for the user demand feature sample, the f related to demand clarity k (x j ) corresponding to α k The weight will be relatively large; for the intention feature sample, Figure 1 Consistency-related f k (x j ) corresponding to α k The weight will be given more importance.

[0184] f k (x j ) is a series of characteristic functions, as follows:

[0185] For the historical behavior data sample h j , assuming it contains the behavior timestamp t j 、Behavior type type j 、Number of goods involved in the behavior j Then f1(h j ) can be a function that determines whether the behavior timestamp is within a reasonable time range, for example where t min and t max It is a reasonable time range set according to business logic; f2(h j) can be a function to determine whether the behavior is complete, such as whether the purchase behavior contains payment success information, etc. Set the complete behavior judgment condition as condition complete ,but

[0186] For user demand feature sample d j , assuming that the demand characteristics are represented by vector in represents the demand for commodity prices, Indicates the demand for product functions, etc. f1(d j ) can be a function to determine whether the price demand is reasonable, for example where p min and p max It is a reasonable price range set according to the commodity market price range; f2(d j ) can be a function to determine whether the demand characteristics are clear. For example, the clarity of the demand vector can be determined by calculating its entropy. Let Entropy(d j ) is a function for calculating the entropy of the demand vector. If the entropy is less than a certain threshold ∈, it means that the demand is clear.

[0187] For the intent feature sample i j , assuming that the intent feature is represented by discrete categories, such as purchase intent buy , browsing intent browse , Collection intent collect wait. f1(i j ) can be a function to determine whether the intentions are consistent, for example, to determine whether there are conflicting intentions at the same time. Let the conflicting intention judgment condition be condition conflict ,but

[0188] If S(x j ) is greater than a certain threshold τ, then the sample is a valid sample and is put into the valid sample set X valid middle.

[0189] After cleaning, we get the valid historical behavior data sample set H valid , effective user demand feature sample set D valid And the effective intention feature sample set I valid .

[0190] 2. Feature Extraction

[0191] 2.1 User Behavior Feature Extraction

[0192] Define the user behavior feature extraction function ExtractBehaviorFeatures(Hvalid ).

[0193] Let h m ∈H valid , assuming it includes the number of purchases n buy,m , browsing time t browse,m , purchase interval Δt m This application combines these sub-features into a user behavior feature vector b in the following way. m :

[0194]

[0195] in:

[0196] β1, β2, and β3 are weighting parameters related to the number of purchases, browsing time, and purchase interval, respectively. In the product recommendation scenario, the number of purchases is more important for characterizing the user's preference for the product, so β1 has a relatively high weighting; browsing time reflects the user's attention to the product, so β2 has a moderate weighting; and the purchase interval reflects the user's purchase frequency, so β3 has a relatively low weighting.

[0197] By dividing the maximum value of each sub-feature in the valid sample set, each sub-feature is normalized so that it can be compared and combined on the same scale.

[0198] This represents the normalized covariance between the number of purchases and the interval between purchases, capturing the correlation between these two behavioral sub-features. For example, a negative covariance indicates that a higher number of purchases is associated with a shorter interval between purchases, reflecting a high frequency of user demand for this type of product.

[0199] The user behavior feature vectors corresponding to all valid historical behavior data samples constitute the user behavior feature set

[0200] 2.2 User Demand Feature Extraction

[0201] Define the user demand feature extraction function ExtractDemandFeatures(D valid ).

[0202] Assume d n ∈D valid , each demand feature can be represented by a vector. Assuming that the dimension of the demand feature vector is r, that is in represents the demand for commodity prices, Indicates the demand for product functions, Indicates demand for product brands, etc.

[0203] This application extracts user demand features through the following methods:

[0204]

[0205] Where: γ k is a weight parameter related to the demand feature dimension k. In the product recommendation scenario, if price has a greater impact on user decision-making, the price-related γ1 weight will be higher; if the user pays more attention to the product function, the γ2 weight will increase accordingly. ReLU(x)=max(0,x) is a rectified linear unit function used to highlight positive demand feature values ​​and filter out noise or unreasonable negative values. For example, the user's demand for product prices is usually positive. The ReLU function can ensure that only reasonable price demand features are retained. μ k It is the mean of the demand feature dimension k. By subtracting the mean, the demand feature can be normalized. For example, for the price dimension, μ1 is the average price demand in all valid samples. For the function dimension, μ2 is some quantitative average value of the function demand in all valid samples (such as the average value of the number of functions, etc.). k Is the unit vector of dimension k, used to map the demand feature value to the corresponding dimension direction. For example, e1 is the unit vector of the price dimension, and e2 is the unit vector of the function dimension. By multiplying with the unit vector, the processed demand feature value is accurately placed on the corresponding dimension. In this way, the user demand feature vector f is obtained d .

[0206] 2.3 User Intent Feature Extraction

[0207] Define the user intention feature extraction function ExtractIntentionFeatures(I valid ). Assume i o ∈I valid , intent features can be represented by discrete categories, such as purchase intent buy , browsing intent browse , Collection intent collect There are s types of intent in total. This application extracts features by constructing an intent feature matrix M.

[0208]

[0209] Then the user intention feature vector f is obtained through matrix operation i :f i =W·vec(M)+b, where:

[0210] W is the weight matrix, with dimension (l is the dimension of the user intent feature vector) and is used to map the vectorized representation vec(M) of the intent feature matrix M to the user intent feature vector space. In product recommendation scenarios, different intent types have different impacts on recommendation strategies, and the weight matrix W can learn these differences. For example, purchase intent places a higher level of urgency on recommendations, so the weights of elements related to purchase intent in the weight matrix W will be relatively large to highlight the impact of purchase intent on the user intent feature vector.

[0211] b is a bias vector with a dimension of l×1, which is used to adjust the overall offset of the user intention feature vector.

[0212] 3. Quantify and classify to generate user portraits

[0213] The user behavior feature set B and the user demand feature vector f d and user intention feature vector f i Perform fusion to obtain a comprehensive feature vector F:

[0214] where δ m is the user behavior feature vector b m The relevant weight parameters are used to adjust the contribution of different user behavior features in the comprehensive feature vector. For example, if the recent user behavior has a greater impact on the user profile, then the δ corresponding to the recent behavior m The weight will be higher.

[0215] Then, the comprehensive eigenvector F is quantized. This application defines a quantization function Quantize(F) that maps continuous eigenvalues ​​to discrete intervals in the following way:

[0216]

[0217] where Δ k is the quantization interval of the kth feature dimension, which is set according to the feature value range and quantization accuracy requirements; round(x) is the rounding function.

[0218] Finally, through the classification function Classify(F quantized ) classifies the quantized feature vectors to generate user profiles. Assuming there are c user profile categories, the classification function calculates the probability of each category based on a classification model (such as a neural network classifier):

[0219]

[0220] Where: W c It is the weight matrix of the classification model, with a dimension of c×n (n is the dimension of the quantized feature vector), which is used to map the quantized feature vector to the category probability space.c is the bias vector with a dimension of c × 1. q is the query vector with a dimension of 1 × c, which is used to calculate the class probability.

[0221] To this end, based on the above solution, in the specific application scenario of generating user portraits in the product recommendation system, the above innovation is compared with the traditional technical processing method, and the significant technical benefits brought are as follows.

[0222] 1. Data cleaning and conversion

[0223] Traditional data cleaning methods are often based on simple rules, such as removing samples with excessive missing values ​​or containing specific invalid identifiers. These methods lack an assessment of the overall validity of the samples. In product recommendation scenarios, this approach can mistakenly delete samples that may have minor flaws but are still valuable overall, or retain samples that appear complete but don't conform to business logic. For example, sample validity can be judged solely based on whether there are missing price values ​​in purchase records, ignoring important factors such as the timing of purchases and their correlation with other behaviors.

[0224] In this application, the effectiveness score is calculated Comprehensively consider multiple characteristic functions f k (x j ) and the corresponding weight α k To assess sample validity. Taking historical behavioral data samples as an example, this not only considers individual factors like behavior timestamps and behavior integrity, but also adjusts the relative importance of each factor through weighting, enabling a more comprehensive and accurate identification of invalid samples. In product recommendation scenarios, this ensures higher-quality data used to generate user profiles, reduces profile bias caused by invalid data, and provides a solid and reliable data foundation for subsequent analysis.

[0225] 2. User behavior feature extraction

[0226] Traditional methods often simply extract single behavioral characteristics, such as focusing solely on the number of purchases or browsing duration, failing to fully explore the potential relationships between behavioral characteristics. Even when considering multiple characteristics, they often simply list them or perform basic statistical calculations, failing to effectively capture the complex interactions between them. For example, when analyzing user purchasing behavior, the inherent connection between the number of purchases and the interval between purchases is not considered, missing the crucial impact of changes in purchase frequency on user profiles.

[0227] In this application, not only are multiple behavioral sub-features such as the number of purchases, browsing time, and purchase interval normalized to make them comparable on the same scale, but the correlation between features is also captured by calculating covariance, e.g. At the same time, the contribution of each sub-feature to the feature vector is adjusted using different weights β1, β2, and β3. In product recommendation scenarios, this approach can more deeply characterize user behavior patterns, for example, identifying the high demand for certain products among users who frequently purchase with short intervals between purchases. This provides richer and more accurate user behavior information for precise recommendations, and more comprehensively reflects user behavior characteristics than traditional methods.

[0228] 3. User demand feature extraction

[0229] Traditional approaches rely solely on simple keyword matching or shallow semantic analysis of user needs, failing to effectively address complex demand expressions or quantify and normalize demand characteristics. For example, when it comes to user demands for product price and functionality, it's difficult to accurately measure the relative importance of different demand dimensions, nor is it possible to properly standardize demand feature values. Consequently, user profiles generated fail to accurately reflect users' true needs.

[0230] In this application, through Extract user demand features and use weight γ k Highlight the importance of different demand dimensions. For example, in product recommendations, the weights can be adjusted according to the user's sensitivity to price, function, brand, etc. At the same time, the ReLU function is used to filter unreasonable negative demands, and the mean μ is subtracted to filter the negative demand. k Normalization is performed to integrate different demand characteristics on the same scale. This approach can more accurately quantify and represent user needs, accurately reflecting the intensity of user demand and preferences for different product attributes, providing more precise user demand information for product recommendation systems. Compared with traditional technologies, it can better capture the diverse and personalized needs of users.

[0231] 4. User Intent Feature Extraction

[0232] Traditional methods for extracting user intent are relatively simplistic, judging intent based on a single action (e.g., clicking a buy button is considered purchase intent). These methods are unable to handle complex multi-intent scenarios and the interrelationships between intents. In product recommendation scenarios, users have multiple intents simultaneously, such as browsing, adding to favorites, and purchasing. Traditional methods struggle to fully and accurately identify and quantify these intents.

[0233] In this application, by constructing the intention feature matrix M and performing matrix operation f i=W·vec(M)+b to extract user intent features. This approach can consider multiple intent types and their distribution across different samples. The weight matrix W can learn the varying degrees of influence of different intents on user profiles, while the bias vector b adjusts the overall offset. In product recommendation scenarios, this can more comprehensively and accurately quantify multiple user intents and their interrelationships. For example, it can distinguish users with strong purchasing intent but who are also cautiously comparing (as evidenced by frequent browsing and saving). This provides more detailed and accurate intent information for the formulation of recommendation strategies, and can provide a deeper understanding of user intent than traditional technologies.

[0234] 5. Quantify and classify to generate user portraits

[0235] Traditional quantification and classification methods often use fixed, simple thresholds or rules that are unable to adapt to the complex and changing nature of user profile data. For example, when quantifying user characteristics into profile categories, they are based on simple, preset numerical ranges and cannot be dynamically adjusted based on the actual data distribution and business needs. Classification models are also overly simplistic, making it difficult to accurately distinguish between different user profile categories.

[0236] In this application, the dynamic quantization interval Δ k and probability calculation based on neural network classifier Quantification and classification are performed. Quantization intervals can be flexibly adjusted based on the distribution of feature data and business needs, improving quantization accuracy. Neural network classifiers can learn complex feature patterns, accurately calculate the probabilities of different categories, and select the category with the highest probability as the user profile category. Compared to traditional methods, they can more accurately map user features to appropriate profile categories, generating profiles that better reflect the user's true characteristics, thereby providing a more precise basis for product recommendations.

[0237] Optionally, the product recommendation module performs the following steps when generating recommendation results:

[0238] Obtain the product features of each product in the product feature library to calculate the semantic similarity between them and the user features in the user profile;

[0239] Filtering a number of candidate products from a product feature library based on the semantic similarity;

[0240] The candidate products are sorted according to the set confidence rules to generate recommendation results.

[0241] Preferably, the above solution is described below in an alternative or preferred manner.

[0242] 1. Calculate semantic similarity

[0243] Assume that the product feature database is where g iRepresents the feature vector of the i-th product, with dimension m, i.e. g i =[g i1 ,g i2 ,…,g im ], each dimension corresponds to different characteristics of the product, such as price, material, function, etc. The feature vector of the user portrait is u=[u1,u2,…,u m ].

[0244] In order to calculate the semantic similarity between product features and user profile features, this application defines a similarity calculation function Sim(g i ,u).

[0245] First, a multi-layer perceptron (MLP) is used to transform the product feature vectors and user profile feature vectors. Assume the MLP consists of three layers: the first layer's weight matrix is ​​W1 (dimension m×h1), and the bias is b1 (dimension h1); the second layer's weight matrix is ​​W2 (dimension h1×h2), and the bias is b2 (dimension h2); and the third layer's weight matrix is ​​W3 (dimension h2×1), and the bias is b3 (dimension 1). Here, h1 and h2 are the number of neurons in the middle layer and can be adjusted based on the experiment and data size. In the product recommendation scenario, multi-layer transformation can be used to explore the complex underlying relationships between product features and user profile features.

[0246] For the product feature vector g i and the user portrait feature vector u, concatenating them together to form a new vector v i =[g i ;u] (dimension is 2m).

[0247] After the first layer transformation: z1=ReLU(W1 T v i +b1), where ReLU(x) = max(0,x) is the rectified linear unit activation function. It introduces nonlinear factors and helps the model learn more complex relationships. In product recommendations, the relationship between different product features and user needs is often not a simple linear relationship. The ReLU function helps capture these complex relationships.

[0248] After the second layer transformation:

[0249] Finally, the semantic similarity score is obtained after the third layer of transformation: The higher the score, the higher the semantic similarity between the product and the user profile.

[0250] 2. Filter alternative products

[0251] Based on the calculated semantic similarity, this application sets a screening rule to determine the candidate products. Let the mean of the semantic similarity be The standard deviation is

[0252] This application selects a threshold Here, k is an adjustable coefficient (for example, k = 0.5). In the product recommendation scenario, if the k value is large, the selected candidate products will be more similar to the user profile, and the recommended results will be more accurate but smaller in number. If the k value is small, the number of candidate products will increase, but some products with slightly lower similarity to the user profile will be included.

[0253] Filter the semantic similarity Sim(g i ,u)≥τ are selected as alternative commodities, and the equipment selection commodity set is where s j is the jth alternative product.

[0254] 3. Sort the candidate products according to the confidence rule

[0255] For each alternative product s j , this application defines a confidence function Confidence(s j ) to sort them. The confidence function takes into account multiple factors, assuming that in addition to the semantic similarity Sim(s j ,u), we also consider the sales volume of the product Sales(s j )、User evaluation score Rating(s j ) and the novelty of the product j ).

[0256] The confidence function is defined as:

[0257] Here, α, β, γ, and δ are weight coefficients, and α + β + γ + δ = 1. In the product recommendation scenario, α represents the importance of semantic similarity in the confidence calculation. If you want the recommendation results to be more closely aligned with the user profile, the value of α should be relatively large. β represents the weight of sales volume. When sales volume is given priority for popular product recommendations, β will be increased. γ represents the weight of user review scores. When product quality and reputation are emphasized, γ can be appropriately increased. δ represents the weight of novelty. If the platform hopes to promote new products, the value of δ will be increased.

[0258] Novelty(s) j ) can be calculated by the difference between the product's listing time and the current time, for example Among them, CurrentTime is the current time, ReleaseTime(sj ) is commodity s j MaxTimeDiff is a normalized parameter, for example, it is set to the longest shelf time of the product on the platform.

[0259] According to the confidence level j ) From high to low, the set of alternative products The final recommendation results are obtained by sorting the products in the list. This recommendation result not only considers the semantic similarity between the product and the user profile, but also integrates other important factors of the product on the platform, which can provide users with more reasonable and diverse product recommendations.

[0260] To this end, in the specific application scenario of product recommendation, the above technology is compared with traditional technical processing methods. The significant benefits brought by the technological innovation are as follows:

[0261] 1. Calculate semantic similarity

[0262] Traditional methods often use simple distance metrics, such as Euclidean distance or cosine similarity, to calculate the semantic similarity between product features and user profiles. These methods measure similarity based solely on the numerical differences in feature vectors and fail to deeply explore the complex semantic relationships between features. For example, when recommending electronic products, a simple distance metric fails to understand the semantic match between the user profile's demand for a "high-performance processor" and the product's "certain processor model," as there is no direct numerical correspondence. Traditional methods ignore the implicit connection between processor performance and model.

[0263] In this application, the above-mentioned deep feature transformation of the product feature vector and the user portrait feature vector is performed by constructing a multi-layer perceptron (MLP). After splicing the two and undergoing multi-layer nonlinear transformation (using ReLU activation function), the complex potential relationship between product features and user portrait features can be excavated. In the electronic product recommendation scenario, MLP can learn the complex mapping between different types of processors and users' demand for high performance, not only based on the surface feature values, but also understand the semantic meaning behind it. This method greatly improves the accuracy of similarity calculation, can more accurately capture the user's potential demand for products, and provide a more reliable basis for subsequent recommendations.

[0264] 2. Filter alternative products

[0265] Traditional screening methods typically use a fixed similarity threshold to select candidate products, failing to consider the overall distribution of the product feature library and the diversity of user profiles. This results in either too many candidate products being selected, resulting in inefficient subsequent sorting and recommendation, or too few, limiting the diversity of recommendations. For example, in a product library with a wide variety of products and widely varying features, a fixed threshold cannot adapt to the matching of different product categories with user profiles, resulting in the omission of some potentially suitable products.

[0266] In this application, the screening threshold is dynamically adjusted according to the mean and standard deviation of semantic similarity. This dynamic adjustment mechanism can adapt to the overall distribution of the product feature library and more flexibly select the appropriate number of candidate products for different user profiles. In product recommendation scenarios, if product features vary significantly within the library, the dynamic threshold can automatically adjust the screening criteria based on the data's discreteness, ensuring that the selected candidate products have both a certain degree of similarity and sufficient diversity. Furthermore, by adjusting the coefficient k, the strictness of the screening can be flexibly controlled based on actual needs, balancing the accuracy and diversity of recommendations.

[0267] 3. Sort the candidate products according to the confidence rule

[0268] Traditional ranking often relies solely on a single factor, such as product sales or user reviews, or simply linearly weights several factors. This approach fails to fully consider the interactions between factors and the varying importance of each factor in different scenarios. For example, when recommending fashion apparel, traditional methods focus solely on sales but ignore users' need for novelty, resulting in recommendations that lack a sense of style and freshness.

[0269] In this application, the new confidence function This approach comprehensively considers multiple factors, including semantic similarity, sales volume, user review scores, and product novelty. Weight coefficients α, β, γ, and δ are used to flexibly adjust the importance of each factor in different scenarios. For example, in a fashion apparel recommendation scenario, if the target user group is trendy, the weight of novelty δ can be appropriately increased to highlight new styles. For a user group focused on quality, the weight of user review score γ can be increased. This multi-factor approach, with flexible weighting, can generate recommendation rankings that better meet user needs based on different user profiles and recommendation scenarios, making it more intelligent and accurate than traditional methods.

[0270] The present application also provides a large-scale model-based intelligent recommendation system, comprising: an electronic device and a backend server. The electronic device is configured with a data collection module and a natural language module. The data collection module is used to collect user behavior data to generate historical behavior data samples; the natural language module is used to process input historical natural language query text based on a natural language processing model to extract user demand feature samples and intent feature samples.

[0271] The backend server is equipped with a user profile generation module, a recommendation object matching module, and a result generation module. The user profile generation module generates a user profile based on historical behavior data samples, user demand feature samples, and intention feature samples; the recommendation object matching module is used to match the user profile with the recommendation object feature library to generate recommendation results; and the result generation module is used to generate push content based on the recommendation results and push it to the user.

[0272] Preferably, the above solution is described below in an alternative or preferred manner.

[0273] 1. Electronic devices

[0274] 1.1 Data Collection Module

[0275] Assume that the user's behavior sequence on the electronic device is B = {b1, b2, ..., b T}, where b t Represents the user's behavior at time t. Behavior b t Can be a tuple, such as b t =(a t ,o t ,t), where a t Indicates the type of behavior (such as click, browse, purchase, etc.), o t Indicates the behavior object (such as product ID, page URL, etc.), and t is the time when the behavior occurs.

[0276] The data collection module generates historical behavior data samples in the following way. Define a behavior filtering function F(b t ), used to determine whether the behavior is valid, for example:

[0277]

[0278] ValidActions is a collection of valid action types, and ValidObjects is a collection of valid action objects. In the product recommendation scenario, ValidActions includes actions such as "Click on the product details page," "Add to cart," and "Purchase product." ValidObjects is a collection of product IDs listed on the platform.

[0279] The historical behavior data sample H is: H={b t |F(b t )=1,t=1,2,...,T}

[0280] 1.2 Natural Language Module

[0281] Assume that the input historical natural language query text is Q = [w1,w2,...,w n ], where w i is the i-th word in the text. Natural language processing models are based on deep learning architectures, such as the Transformer architecture.

[0282] Define word embedding function E(w i ), each word w i Mapped to a d-dimensional vector e i :

[0283] e i =Embedding(w i ), where Embedding is the pre-trained word embedding matrix with dimension |V|×d, and |V| is the vocabulary size.

[0284] After word embedding, we get the word vector sequence E=[e1,e2,...,e n ]. Input it into the Transformer model, and the output of the Transformer model is Z=[z1,z2,...,z n ], where z i It is a vector representation processed by multi-layer self-attention mechanism and feedforward neural network.

[0285] Define the user demand feature extraction function ExtractDemand(Z): where α i is the weight coefficient, which is calculated by the attention mechanism:

[0286] Here q D It is a query vector extracted from demand features and is learned through training. In the product recommendation scenario, it can capture key information related to user needs.

[0287] Similarly, define the intent feature extraction function ExtractIntention(Z): where q I is the query vector for intent feature extraction.

[0288] 2. Backend Server

[0289] 2.1 User portrait generation module

[0290] Assume that the behavior type set contained in the historical behavior data sample H is The behavior object collection is For each behavior type and behavior objects Count its frequency of occurrence f(a,o): in is an indicator function that takes on the value 1 if the condition in the brackets is true, and 0 otherwise.

[0291] User behavior feature vector U B It can be expressed as:

[0292] Combine the user demand feature sample D and the intention feature sample I obtained from the natural language module to generate the user profile U in the following way: U = γ1·U B +γ2·D+γ3·I,

[0293] γ1, γ2, and γ3 are weight coefficients, obtained through model training, used to adjust the contribution of different features to the user profile. In the product recommendation scenario, if the user's historical behavior is more important for portraying the user profile, the value of γ1 is relatively large.

[0294] 2.2 Recommended Object Matching Module

[0295] Assume that the recommended object feature library is where r k It is a d′-dimensional vector representing the features of the k-th recommended object (such as a product).

[0296] Define a similarity measurement function based on kernel function Similarity(U,r k ):

[0297] Here, σ is the bandwidth parameter of the kernel function, which controls the decay rate of similarity. In the product recommendation scenario, a smaller σ means greater sensitivity to feature differences, and only products with very similar features are considered highly similar.

[0298] The recommendation result R is a list of recommended objects sorted from high to low by similarity:

[0299]

[0300] 2.3 Result Generation Module

[0301] Set the recommendation result in It is the recommended object after sorting.

[0302] Define a content generation function Generate push content based on the characteristics of the recommended object. For example, for product recommendations, the push content includes information such as the product title, description, and image link.

[0303] The final push content set generated is C = [Content1, Content2, ..., Content s ] and push it to the user.

[0304] The above embodiments are only used to illustrate the embodiments of the present invention, and are not intended to limit the embodiments of the present invention. Ordinary technicians in the relevant technical field can make various changes and modifications without departing from the spirit and scope of the embodiments of the present invention. Therefore, all equivalent technical solutions also fall within the scope of the embodiments of the present invention, and the scope of patent protection of the embodiments of the present invention should be defined by the claims. The systems, devices, modules or units described in the above embodiments are specifically implemented by computer chips or entities, or by products with certain functions.

Claims

1. A commodity intelligent recommendation system based on a large model, characterized by: include: An electronic device and a backend server, wherein the electronic device is configured with a data collection module and a natural language module, wherein the data collection module is used to collect user behavior data to generate historical behavior data samples, and the natural language module is used to process input historical natural language query texts based on a natural language processing model to extract user demand feature samples and intention feature samples; The background server is configured with a user portrait generation module, a product recommendation module, and the result generation module. The user portrait generation module generates a user portrait based on historical behavior data samples, user demand feature samples, and intention feature samples; the product recommendation module is used to match the user portrait with the product feature library to generate a recommendation result. The result generation module is used to generate multi-screen push content based on the recommendation result and refer to the display size of the electronic device and push it to the user one by one. At the same time, it collects user feedback data on the current one-screen push content, so that the product recommendation module optimizes the multi-screen push content accordingly and pushes it again.

2. The system according to claim 1, wherein: The data collection module is also used to collect historical recommendation content data, the product recommendation module is also used to adjust the recommendation results based on the historical recommendation content data, and the result generation module is used to generate multi-screen push content based on the adjusted recommendation results and with reference to the display size of the electronic device.

3. The system according to claim 2, characterized in that When collecting user behavior data and / or historical recommended content data, the data collection module detects the user's behavior records on web pages and APPs based on the set crawler component to collect user behavior data and / or historical recommended content data and store them in the basic database according to the set storage mechanism.

4. The system according to claim 2, wherein: The data collection module is deployed with a data collection framework to trigger the collection of user behavior data and / or historical recommended content data based on the data collection framework.

5. The system according to claim 2, wherein: The backend server is also configured with a model training module for performing forward propagation and backpropagation calculations on model parameters of the natural language processing model to be trained based on the natural language query text sample until the set model training end condition is met, so as to obtain the trained natural language processing model; When the natural language processing model processes the input historical natural language query text, it extracts user demand feature samples and intention feature samples therefrom through logical reasoning.

6. The system according to claim 2, wherein: The natural language processing model is a natural language processing component. When the natural language processing model processes the input historical natural language query text, it extracts user demand feature samples and intention feature samples from it through text semantic analysis. The text semantic analysis includes at least one of word segmentation, part-of-speech tagging, named entity recognition, and syntactic analysis.

7. The system according to claim 1, wherein: When generating a user portrait, the user portrait generation module performs the following steps: Cleaning and converting the historical behavior data samples, user demand feature samples, and intention feature samples, removing invalid samples therefrom, and obtaining valid historical behavior data samples, valid user demand feature samples, and valid intention feature samples; Extracting features from the valid historical behavior data samples, the valid user demand feature samples, and the valid intention feature samples to obtain user behavior features, user demand features, and user intention features; The user behavior characteristics, user demand characteristics, and user intention characteristics are quantified and classified to generate a user profile.

8. The system according to claim 1, wherein: The product recommendation module performs the following steps when generating recommendation results: Obtain the product features of each product in the product feature library to calculate the semantic similarity between them and the user features in the user profile; Filtering a number of candidate products from a product feature library based on the semantic similarity; The candidate products are sorted according to the set confidence rules to generate recommendation results.

9. An intelligent recommendation system based on a large model, characterized in that: include: Electronic device and backend server. The electronic device is equipped with a data collection module and a natural language module. The data collection module is used to collect user behavior data to generate historical behavior data samples; the natural language module is used to process the input historical natural language query text based on a natural language processing model to extract user demand feature samples and intention feature samples. The backend server is equipped with a user profile generation module, a recommendation object matching module, and a result generation module. The user profile generation module generates a user profile based on historical behavior data samples, user demand feature samples, and intention feature samples; the recommendation object matching module is used to match the user profile with the recommendation object feature library to generate recommendation results; and the result generation module is used to generate push content based on the recommendation results and push it to the user.

Images