E-commerce intelligent recommendation system based on user behavior big data

Abstract

The present application relates to the technical field of e-commerce intelligent recommendation, in particular to an e-commerce intelligent recommendation system based on user behavior big data, which comprises a data acquisition and storage unit, a data preprocessing unit, a model construction and training unit and a recommendation generation and output unit; in the present application, the data acquisition and storage unit is deployed with a page tracking point on each page of an e-commerce platform, collects data such as user behavior, device, geographic location and access time and stores them in a distributed database; the data preprocessing unit extracts data, constructs a user behavior sequence and encodes product and behavior features; the model construction and training unit obtains a user interest vector by using a long short-term memory network combined with an attention mechanism and updates model parameters by means of a deep Q network; the recommendation generation and output unit filters products, sorts and displays them according to the interest vector, collects feedback to optimize the system, improves recommendation accuracy, enhances user shopping experience and helps the e-commerce platform improve operation efficiency and product sales conversion rate.

Description

Technical Field

[0001] The present invention relates to the technical field of e-commerce intelligent recommendation, and in particular to an e-commerce intelligent recommendation system based on user behavior big data. Background Art

[0002] E-commerce intelligent recommendation is an important technology. With the vigorous development of e-commerce, the number of products on e-commerce platforms has exploded, making it increasingly difficult for users to select the items they like from the massive amount of products. At the same time, e-commerce companies are also faced with the challenge of how to improve user retention and promote product sales.

[0003] In today's e-commerce shopping scenarios, users often face the problem of information overload when purchasing products. Traditional recommendation systems, due to their overly general data collection methods, fail to distinguish the importance of different behaviors, resulting in insufficient processing of user behavior data and insufficient analytical and mining capabilities. It is difficult to accurately grasp users' specific preferences for sports shoe brands, styles, functions, and prices, and it is also impossible to recommend suitable products based on market popularity. It is often the case that recommended products are neither what users want nor currently popular, making it difficult for users to sift through a large number of products and reducing consumers' willingness to buy. To solve this technical problem, we provide an e-commerce intelligent recommendation system based on user behavior big data. Summary of the Invention

[0004] The purpose of the present invention is to provide an e-commerce intelligent recommendation system based on user behavior big data to solve the problems raised in the above background technology.

[0005] To achieve the above objectives, an e-commerce intelligent recommendation system based on user behavior big data is provided, which includes a data acquisition and storage unit, a data preprocessing unit, a model building and training unit, and a recommendation generation and output unit;

[0006] The data collection and storage unit deploys embedded points on each page of the e-commerce platform, which are used to collect user behavior data as well as user device information, geographic location and access time, and store the collected behavior data in a distributed database;

[0007] The data preprocessing unit extracts and processes the collected behavior data from the distributed database, then arranges the user behavior data in chronological order to construct a user behavior sequence, where each sequence contains multiple behavior nodes, each node contains product features and behavior features, and finally encodes the product features and behavior features and converts them into numerical vectors;

[0008] The model construction training unit introduces a long short-term memory network as a core model, takes the encoding vector of the obtained user behavior sequence after preprocessing as input, performs feature extraction and learning through a plurality of long short-term memory network units, outputs the hidden state of each time step, applies an attention mechanism on the hidden state, calculates the attention weight of each time step, obtains the final user interest representation vector through weighted summation, defines the state, action and reward, adopts a deep Q network as a reinforcement learning algorithm, selects the optimal recommendation action according to the current state, and updates the model parameters according to the reward signal;

[0009] The recommendation generation output unit filters out the goods related to the user interest from the goods library according to the user interest representation vector, forms a candidate goods set, and outputs a recommendation strategy, sorts the candidate goods set, determines the order of the recommended goods, and finally displays the sorted recommended goods list to the user, and records the feedback of the user to the recommended result.

[0010] As a further improvement of the technical solution, in the data acquisition and storage unit, an adaptive sampling strategy is adopted when collecting user behavior data, and the specific steps are as follows:

[0011] Define a set of behavior data importance indicators, and the behavior data importance indicators are the correlation between behavior and purchase conversion, the rarity of behavior and the timeliness of behavior, and each indicator is assigned an initial weight;

[0012] For each user behavior event, calculate its importance score according to the behavior data importance indicators and their initial weights, and set a dynamic sampling rate according to the importance score, and perform adaptive sampling through the dynamic sampling rate to obtain user behavior data.

[0013] As a further improvement of the technical solution, the data preprocessing unit, when constructing the user behavior sequence, has the following specific method for solving the problem of behavior sequence interruption caused by long-time inaction:

[0014] Set a time interval threshold, and when it is found that the time interval between two adjacent behavior nodes in the user behavior sequence is greater than the time interval threshold, it is determined that the sequence is interrupted;

[0015] For the interrupted sequence, a sequence filling strategy is adopted, that is, the behavior characteristics before interruption are analyzed, if the previous sequence of behaviors is concentrated on browsing a certain category of goods, several similar goods are selected from the user's historical browsing, and inserted into the interruption place according to the browsing probability to simulate the user's subsequent behavior.

[0016] As a further improvement of the technical solution, the data preprocessing unit adopts a knowledge graph-based encoding method when encoding the goods features and behavior features, and the specific steps are as follows:

[0017] The e-commerce field knowledge graph is constructed, nodes include commodities, brands, categories and attributes, and edges represent the association relationship therebetween, for the commodity features, the embedding vectors of the corresponding category nodes, brand nodes and key attribute nodes of the commodity are found according to the knowledge graph, and the embedding vectors are spliced to serve as the encoding vectors of the commodity features;

[0018] For the behavior features, the one-hot encoding is used to map the behavior features into corresponding behavior vectors, and then the behavior vectors are fused with the commodity feature encoding vectors of the behavior nodes to obtain the final encoding vectors containing the commodity and behavior information.

[0019] As a further improvement of the technical solution, the model construction and training unit comprises an extraction learning module, and when the multi-layer long short-term memory network unit in the extraction learning module performs feature extraction and learning, a gating mechanism is introduced, and the specific operation is as follows:

[0020] On the basis of the input gate, the forget gate and the output gate of the long short-term memory network, an adaptive adjustment gate is added, and for each time step in the input sequence, the input of each gate is obtained through linear transformation;

[0021] The input of the adaptive adjustment gate is the dynamic adjustment of the feature extraction according to the current input and the previous state.

[0022] As a further improvement of the technical solution, the model construction and training unit comprises an attention mechanism module, and when the attention mechanism module applies the attention mechanism to the hidden state, a multi-head attention mechanism is adopted, and the specific calculation method is as follows:

[0023] A hidden state sequence is set, a hidden state vector is extracted from the hidden state sequence, and the hidden state vector is mapped to multiple subspaces through linear transformation to generate query vectors, key vectors and value vectors of each head;

[0024] For each head, the attention score is calculated, the attention weight is obtained through the softmax function, the context vector of each head is calculated, and finally The context vectors of the heads are spliced and linearly transformed to obtain the final attention output.

[0025] As a further improvement of the technical solution, the model construction and training unit comprises a reinforcement learning module, and when the reinforcement learning module adopts a deep Q network as a reinforcement learning algorithm, the specific method of defining the state, action and reward is as follows:

[0026] The state is composed of the interest representation vector of the current user, the information of the e-commerce page where the user is currently located, and the historical behavior statistical features of the user;

[0027] Select a certain number of commodities from the commodity library as recommended candidate actions, and each action corresponds to a candidate commodity set;

[0028] If the user performs a purchase behavior within a certain time after seeing the recommended commodity, a positive reward is given, if the user clicks on the recommended commodity to enter the detail page, half of the positive reward is given, and if the user has no reaction to the recommended commodity, a negative reward is given.

[0029] As a further improvement of the technical solution, the recommendation generation output unit adopts a sorting algorithm that fuses user preferences and commodity popularity when sorting the candidate commodity set, and the specific steps are as follows:

[0030] For each candidate commodity, calculate its user preference score, extract the user's preference degree for the category and brand to which the commodity belongs from the user behavior data, and obtain the user preference score by weighted summation;

[0031] Calculate the commodity popularity score, count the number of views and purchases of the commodity within a certain time, and aggregate according to the preset weight, and finally calculate the sorting score of the candidate commodity according to the user preference score and the commodity popularity score;

[0032] Collect user feedback on the recommendation results, organize these feedback data into structured data form, and analyze the feedback data using association rule mining algorithm, and optimize the recommendation system units according to the analysis results.

[0033] Compared with the prior art, the beneficial effects of the present application are:

[0034] In the e-commerce intelligent recommendation system based on user behavior big data, by constructing user behavior sequence, the interruption problem caused by long time non-operation can be effectively handled, by analyzing the characteristics of the previous sequence to fill in the reasonable sequence, the coherence and accuracy of the behavior analysis are guaranteed, based on the coding method of knowledge graph, the potential relationship between goods and behavior characteristics is deeply mined, the feature coding more accurately reflects the user preference, the model construction training unit introduces long short-term memory network, gate mechanism and multi-head attention mechanism, which can more effectively capture the dynamic changes of user interest, the reinforcement learning module defines the state, action and reward reasonably, the recommendation strategy is continuously optimized, the recommendation generation output unit fuses user preferences and commodity popularity to sort commodities, and optimizes the system according to user feedback, improves the accuracy and effectiveness of the recommendation, improves the user shopping experience, and improves the commodity sales conversion rate and operation efficiency of the e-commerce platform. BRIEF DESCRIPTION OF DRAWINGS

[0035] Figure 1 The overall block diagram of the present application.

[0036] The meanings of the various labels in the figure are:

[0037] 1. Data acquisition and storage unit; 2. Data preprocessing unit; 3. Model construction and training unit; 31. Extraction learning module; 32. Attention mechanism module; 33. Reinforcement learning module; 4. Recommendation generation and output unit. DETAILED DESCRIPTION

[0038] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0039] This invention provides an e-commerce intelligent recommendation system based on user behavior big data, please refer to Figure 1 As shown, it includes a data acquisition and storage unit 1, a data preprocessing unit 2, a model building and training unit 3, and a recommendation generation and output unit 4;

[0040] The data collection and storage unit 1 deploys embedded points on each page of the e-commerce platform, which are used to collect user behavior data as well as user device information, geographic location and access time, and store the collected behavior data in a distributed database.

[0041] In the data collection and storage unit 1, an adaptive sampling strategy is used to collect user behavior data. The specific steps are as follows:

[0042] Define a set of behavioral data importance indicators ,in The correlation between behavior and purchase conversion is measured by counting the proportion of users who purchased goods within a certain period of time after a certain behavior in historical data. is the rarity of the behavior, that is, the frequency of the behavior in all user behaviors, For the timeliness of the behavior, the closer the behavior is to the current time, the higher its importance is. Assign an initial weight to each indicator ,and .

[0043] For each user behavior event , calculate its importance score ; Set dynamic sampling rate based on importance score and set high importance threshold and the importance threshold ,when When the sampling rate is ,when When the sampling rate is ,when When the sampling rate is By this adaptive sampling, both the key behavior data can be collected and the unnecessary data storage can be reduced.

[0044] The data preprocessing unit 2 extracts the collected behavior data from the distributed database and processes it, arranges the user's behavior data in time sequence, constructs the user behavior sequence, wherein each sequence contains multiple behavior nodes, each node contains commodity features and behavior features, and finally encodes the commodity features and behavior features and converts them into numerical vectors.

[0045] The specific method of the data preprocessing unit 2 for handling the problem of behavior sequence interruption caused by long-time inaction when constructing the user behavior sequence is as follows:

[0046] In the user behavior sequence, long-time inaction may cause logical discontinuity in the sequence, which will affect the learning and understanding of user behavior patterns by the subsequent model. By setting a time interval threshold , the sequence interruption can be clearly defined, providing a clear judgment standard for subsequent processing.

[0047] For the user behavior sequence , wherein represents the th behavior node, each behavior node contains the time when the behavior occurs and , if the two adjacent behavior nodes satisfy , it is determined that the sequence is interrupted. By setting a clear threshold, the determination of sequence interruption is objective and operable, avoiding the uncertainty of subjective judgment.

[0048] For the sequence determined to be interrupted, the behavior nodes before the interruption are analyzed, and the product categories involved in these behavior nodes are counted. Based on the user's historical behavior characteristics, the user's historical preference information can be fully utilized to make the filled behavior more in line with the user's actual needs.

[0049] Let the set of similar products browsed by the user in history be , for each product , calculate its browsing probability . The browsing probability can be calculated by the ratio of the historical browsing times of the product to the total browsing times of similar products, making the selection of the filled product more scientific and reasonable, and avoiding the irrationality that may be caused by random selection.

[0050] According to the calculated browsing probability , select products from the product set, and select The commodities are inserted into the interruption according to the browsing probability from high to low, forming a new behavior sequence, so that the filled behavior sequence is closer to the real situation, effectively solves the interruption problem of the behavior sequence, keeps the behavior sequence coherent, provides more complete and accurate data for subsequent model learning, and helps improve the prediction and recommendation ability of the model on user behavior.

[0051] When encoding the commodity features and behavior features, the data preprocessing unit 2 adopts a knowledge graph-based encoding method, and the specific steps are as follows:

[0052] In the e-commerce field, there are complex association relationships among commodities, brands, categories, and attributes. Building a knowledge graph can clearly represent these relationships in a graphical manner. The knowledge graph can integrate multiple aspects of information, providing rich semantic information and association basis for subsequent encoding of commodity features and behavior features, which helps to more comprehensively and accurately understand and process data. Collect commodity data, brand information, category classification, and commodity attribute data on the e-commerce platform. The commodity, brand, category, and attribute are taken as nodes, and the edges represent the association relationship between them. For example, the commodity node and the brand node it belongs to are connected by a "brand association" edge, the commodity node and the category node it belongs to are connected by a "category association" edge, and the commodity node and the attribute node it corresponds to are connected by an "attribute association" edge. In this way, an e-commerce domain knowledge graph is constructed. wherein is a node set, is an edge set. The knowledge graph can intuitively display the complex relationships in the e-commerce field, providing a structured information foundation for the subsequent encoding process, and facilitating the mining of potential relationships between data.

[0053] For a given commodity , find its corresponding category node , brand node , and key attribute node set in the knowledge graph . Based on the knowledge graph, the node can be accurately located to obtain the relevant feature information of the commodity, avoiding the information loss or inaccuracy that may occur in traditional methods. In order to convert the node information in the knowledge graph into a numerical form that can be processed by a computer, the node needs to be embedded. By concatenating the embedding vectors of the category, brand, and attribute nodes, the multiple features of the commodity can be integrated into a vector, forming an encoding vector of the commodity features, which is convenient for subsequent model processing.

[0054] Use a graph embedding algorithm to embed the nodes in the knowledge graph and obtain the embedding vector of each node. Let the embedding vector of the category node be , the embedding vector of the brand node be , attribute node The embedding vector of the attribute node is , the encoding vector of the commodity feature The concatenation of these vectors is, that is, , wherein represents the vector concatenation operation, and the generated commodity feature encoding vector contains rich commodity information, which is conducive to the model better understanding the features of the commodity and improving the performance of the model.

[0055] Define the behavior feature set For a behavior , use one-hot encoding to map it to a vector of length , wherein only the th element is 1, and the remaining elements are 0. One-hot encoding is simple and intuitive, and can clearly represent different behavior categories, and is convenient for fusion with commodity feature encoding vectors, converting behavior features into a computable vector form, providing a basis for subsequent feature fusion.

[0056] For each behavior node, the one-hot encoding vector of the behavior is fused with the commodity feature encoding vector of the behavior node to obtain the final encoding vector containing commodity and behavior information Feature fusion integrates the information of commodities and behaviors, so that the encoding vector can more comprehensively describe the interaction between users and commodities, and provides the model with more rich input features.

[0057] The model construction and training unit 3 introduces a long short-term memory network as the core model, takes the encoding vector of the user behavior sequence obtained after preprocessing as the input, performs feature extraction and learning through multiple long short-term memory network units, outputs the hidden state of each time step, and applies an attention mechanism on the hidden state to calculate the attention weight of each time step. The final user interest representation vector is obtained by weighted summation, and finally the state, action and reward are defined, a deep Q network is used as a reinforcement learning algorithm, the optimal recommendation action is selected according to the current state, and the model parameters are updated according to the reward signal.

[0058] The model construction and training unit 3 includes an extraction and learning module 31. When the multiple long short-term memory network units in the extraction and learning module 31 perform feature extraction and learning, a gating mechanism is introduced, and the specific operation is as follows:

[0059] ​The input gate, the forget gate and the output gate of the traditional long short-term memory network may not dynamically and flexibly adjust the focus of feature extraction according to the input data when processing complex sequence data. The introduction of the adaptive adjustment gate can more finely screen and adjust the information according to the current input and the previous state, thereby better capturing the long-term dependence in the user behavior sequence and improving the feature extraction effect.

[0060] On the basis of the traditional long short-term memory network, an adaptive adjustment gate is added. For each time step in the input sequence, the input vector and the hidden state at the previous time are set, the outputs of the input gate, the forget gate and the output gate are calculated through linear transformation and sigmoid function, wherein the output gate is used to control how much new information can enter the cell state, the forget gate is used to control how much information of the cell state at the previous time needs to be forgotten, and the output gate controls how much information of the cell state is output to the current hidden state. Finally, the output of the adaptive adjustment gate is calculated through linear transformation and hyperbolic tangent function. The adaptive adjustment gate dynamically adjusts the focus of feature extraction according to the current input and the previous state, increases the flexibility and adaptability of the model, and can dynamically adjust the information processing mode according to different input conditions, so that the model is more efficient when processing complex sequence data.

[0061] After introducing the adaptive adjustment gate, the hidden state at the current time needs to be recalculated to fully utilize the information screened and adjusted by the adaptive adjustment gate, so as to more accurately represent the feature information at the current time and provide better input for the subsequent attention mechanism and reinforcement learning.

[0062] Firstly, the forget gate determines how much information of the hidden state at the previous time is retained, the input gate and the adaptive adjustment gate jointly determine how much new information enters the cell state, and the sum is mapped to the interval through the hyperbolic tangent function. Finally, the output gate controls how much information is output to the current hidden state. The new hidden state calculation method comprehensively considers the role of the adaptive adjustment gate, can more effectively integrate historical information and current input information, improves the representation ability of the hidden state, and enables the model to more accurately represent the behavior characteristics of the user at each time step.

[0063] The model construction and training unit 3 includes an attention mechanism module 32. When the attention mechanism module 32 applies the attention mechanism to the hidden state, a multi-head attention mechanism is adopted, and the specific calculation method is as follows:

[0064] The traditional single attention mechanism may not comprehensively capture different feature information in the hidden state sequence. The hidden state vector is mapped to multiple subspaces through linear transformation to generate multiple query vectors, key vectors and value vectors, which can enable the model to pay attention to the hidden state sequence from different angles and subspaces, thereby more comprehensively and meticulously extracting feature information.

[0065] Let the hidden state sequence be , where is the hidden state vector of the th time step, and the number of heads is set to For each head , the query vector , the key vector and the value vector are generated through linear transformation. The multi-head attention mechanism increases the model's expressive power and the diversity of feature extraction through the mapping of multiple subspaces, enabling it to capture more rich feature information in the hidden state sequence.

[0066] For each head , the attention score is calculated, with the formula ; where is the query vector of the th head at the th time step, is the key vector of the th head at the th time step, is the dimension of the key vector, divided by is to prevent the dot product result from being too large, causing the softmax function gradient to disappear, and then the attention weight is obtained through the softmax function. The calculation of the attention score and weight can adaptively determine the importance of each time step in the current head, enabling the model to focus on the more critical parts of the sequence and improving the efficiency and accuracy of information processing.

[0067] For each head , the context vector is calculated, with the formula ; where is the value vector of the th head at the th time step. The context vector integrates the information of the attention weight and the value vector, effectively capturing the key information in the hidden state sequence, and the context vector of each head reflects the features of the sequence from different subspaces and angles.

[0068] The context vectors of the heads are concatenated to obtain the concatenated vector , and then the final attention output is obtained through linear transformation , with the formula ; where is the output weight matrix, the concatenation and linear transformation operations can fuse and integrate the feature information extracted by multiple heads, obtain a comprehensive and more representative feature representation, fully exert the advantages of multi-head attention mechanism, and the final attention output can more comprehensively and accurately represent the user's interest, provide better feature input for subsequent reinforcement learning and commodity recommendation, thereby improving the recommendation accuracy and effect of the e-commerce intelligent recommendation system.

[0069] The model construction and training unit 3 includes a reinforcement learning module 33. When the deep Q network is used as the reinforcement learning algorithm, the specific method of defining the state, action and reward is as follows:

[0070] In reinforcement learning, accurately describing the state of the current environment is crucial for the agent to make reasonable decisions. Combining the current user's interest representation vector, the user's current e-commerce page information, and the user's historical behavior statistical features as the state can comprehensively and meticulously depict the user's current situation, providing rich context information for the recommendation system, so that the agent can make recommendations that are more in line with the user's needs based on this state.

[0071] Let the current user's interest representation vector be , which is obtained through the previous attention mechanism module 32 and can reflect the user's interest. The user's current e-commerce page information can be represented by a vector , and the user's historical behavior statistical features can be represented by a vector , which can include statistical information such as the number of goods browsed, the number of purchases, and the average purchase amount of the user in the past period. Then the state can be represented as , so that the recommendation system can make more accurate product recommendations based on the user's specific state, improving the relevance and accuracy of the recommendations, and thus improving the user's shopping experience on the e-commerce platform.

[0072] Let the product library be , and select products from it to form a candidate product set . Each candidate product set corresponds to a recommendation action , so that the recommendation system can provide users with more diverse and personalized product recommendations, increasing the user's chances of discovering interesting products, and thus improving the user's purchase conversion rate and the platform's sales.

[0073] The reward mechanism is a core part of reinforcement learning, which is used to encourage the agent to take actions that can maximize long-term benefits. Different rewards are given according to different user feedback on recommended products, which can guide the agent to learn which recommendation actions are effective and which are not, thereby continuously optimizing the recommendation strategy.

[0074] Let the positive reward be , if the user makes a purchase within a certain time after seeing the recommended product, give a positive reward , if the user clicks on the recommended product to enter the detail page, give half of the positive reward , if the user has no reaction to the recommended product, give a negative reward This reward setting can reasonably measure different feedback of users to recommended products, encourage the agent to preferentially select recommended actions that can promote user purchase or click, while avoiding making ineffective recommendations, improving the efficiency and effectiveness of the recommendation system.

[0075] Through the guidance of the reward mechanism, the agent can continuously adjust the recommendation strategy to improve the quality of the recommendation, thereby increasing the user's purchase willingness and the platform's revenue, achieving optimization and improvement of the recommendation system.

[0076] The recommendation generation output unit 4 filters out the goods related to the user's interest from the goods library according to the user's interest representation vector, forms a candidate goods set, and outputs the recommendation strategy, sorts the candidate goods set, determines the order of the recommended goods, and finally displays the sorted recommended goods list to the user, while recording the user's feedback on the recommended results.

[0077] When sorting the candidate goods set, the recommendation generation output unit 4 adopts a sorting algorithm that combines user preferences and product popularity, the specific steps are as follows:

[0078] Users on e-commerce platforms often have different degrees of preference for different categories and brands of goods. Understanding the user's preference for the category and brand of the candidate goods can more accurately grasp the user's interest points. The user preference score obtained by weighted summation can consider the influence of category and brand preference on user selection of goods, thereby providing a basis for goods sorting that is more in line with user needs.

[0079] Let the candidate goods be , the category it belongs to be , and the brand it belongs to be , the user's preference degree for the category and the brand are extracted from the user behavior data, and the weights of the category preference and brand preference are respectively and , then the user's preference score for the candidate goods is The calculation formula is , taking into account users' preferences for categories and brands, it can more carefully portray users' personalized needs, avoid the limitation of evaluating user preferences from only a single dimension, and make product sorting more in line with users' actual interests.

[0080] The popularity of a product reflects its popularity in the market. At a certain time The number of views in , the purchase amount is , assign preset weights to page views and purchases and , then the product The popularity score The calculation formula is ;in and are all candidate products at time The maximum number of views and the maximum number of purchases within a product are combined to measure the popularity, which can fully reflect the popularity of the product. The influence of the two on the popularity score can be flexibly adjusted through weight distribution, making the sorting results more reasonable.

[0081] Assign weights to user preference scores and product popularity scores respectively and , then the candidate product Ranking score The calculation formula is ,in It can be adjusted according to user groups, integrating two important factors: user preferences and product popularity, to achieve an organic combination of personalized recommendations and popular recommendations. It can meet the needs of different users in different scenarios and improve the performance and adaptability of the recommendation system.

[0082] Collect user feedback on recommendation results, including purchase behavior, click behavior, dwell time, and skip behavior information. This information is organized according to certain rules to form structured data. An association rule mining algorithm is then used to analyze this structured feedback data, identifying association rules that meet the requirements. The analysis results are then fed back to each unit of the recommendation system, achieving closed-loop optimization of the recommendation system. This allows for continuous adjustment and improvement of recommendation strategies based on actual user feedback, enhancing the adaptability and accuracy of the recommendation system and enabling it to continuously meet user needs.

[0083] In the application, the data acquisition and storage unit 1 is arranged on each page of the e-commerce platform to collect user behavior, device, geographic location and access time data and store them in a distributed database, the data preprocessing unit 2 extracts the data, constructs a user behavior sequence, encodes the goods and behavior characteristics, the model construction and training unit 3 obtains a user interest vector by using a long short-term memory network combined with an attention mechanism, updates the model parameters by means of a deep Q network, the recommendation generation output unit 4 filters goods, sorts and displays according to the interest vector, collects feedback to optimize the system, improves the recommendation accuracy, enhances the user shopping experience, and helps the e-commerce platform to improve the operation efficiency and the goods sales conversion rate.

[0084] The basic principles, main features and advantages of the present application are shown and described above. Those skilled in the art should understand that the present application is not limited to the above-mentioned embodiments, and the above-mentioned embodiments and descriptions in the specification are only preferred examples of the present application and are not intended to limit the present application. Without departing from the spirit and scope of the present application, various changes and improvements can be made to the present application, and these changes and improvements all fall within the scope of the claimed present application. The scope of protection of the present application is defined by the appended claims and their equivalents.

Claims

1. E-commerce intelligent recommendation system based on user behavior big data, characterized by: It includes a data acquisition and storage unit (1), a data pre-processing unit (2), a model building and training unit (3), and a recommendation generation and output unit (4); The data collection and storage unit (1) deploys embedded points on each page of the e-commerce platform, and the embedded points are used to collect user behavior data and user device information, geographic location and access time, and store the collected behavior data in a distributed database; The data preprocessing unit (2) extracts and processes the collected behavior data from the distributed database, and then arranges the user behavior data in chronological order to construct a user behavior sequence, wherein each sequence contains multiple behavior nodes, and each node contains product features and behavior features. Finally, the product features and behavior features are encoded and converted into numerical vectors; The model construction training unit (3) introduces a long short-term memory network as a core model, takes the encoding vector of the user behavior sequence obtained after preprocessing as input, performs feature extraction and learning through multi-layer long short-term memory network units, outputs the hidden state of each time step, applies an attention mechanism to the hidden state, calculates the attention weight of each time step, obtains the final user interest representation vector by weighted summation, and finally defines the state, action and reward, adopts a deep Q network as a reinforcement learning algorithm, selects the optimal recommended action according to the current state, and updates the model parameters according to the reward signal; The model construction training unit (3) includes an attention mechanism module (32). When the attention mechanism module (32) applies the attention mechanism on the hidden state, a multi-head attention mechanism is adopted. The specific calculation method is as follows: Set the hidden state sequence, extract the hidden state vector from the hidden state sequence, and map the hidden state vector to multiple subspaces through linear transformation to generate The query vector, key vector, and value vector of each head; For each head, calculate the attention score, and get the attention weight through the softmax function, and then calculate the context vector of each head. The context vectors of the heads are concatenated and linearly transformed to obtain the final attention output. The recommendation generation output unit (4) selects products related to the user's interests from the product library according to the user's interest representation vector, forms a candidate product set, outputs a recommendation strategy, sorts the candidate product set, determines the order of recommended products, and finally displays the sorted recommended product list to the user, while recording the user's feedback on the recommendation results. When sorting the candidate product set, the recommendation generation output unit (4) adopts a sorting algorithm that integrates user preferences and product popularity. The specific steps are as follows: For each candidate product, calculate its user preference score. Extract the user's preference for the product's category and brand from the user behavior data, and obtain the user preference score through weighted summation. Calculate the popularity score of a product by counting the number of views and purchases of the product within a certain period of time, summarizing them according to the preset weights, and finally calculating the ranking score of the candidate products based on the user preference score and the product popularity score; Collect user feedback on recommendation results, organize the feedback data into structured data, use association rule mining algorithms to analyze the feedback data, and optimize each unit of the recommendation system based on the analysis results.

2. The e-commerce intelligent recommendation system based on user behavior big data according to claim 1 is characterized by: In the data acquisition and storage unit (1), an adaptive sampling strategy is adopted when collecting user behavior data. The specific steps are as follows: Define a set of behavioral data importance indicators, where the behavioral data importance indicators are the correlation between behavior and purchase conversion, the rarity of behavior, and the timeliness of behavior, and assign an initial weight to each indicator; For each user behavior event, its importance score is calculated based on the behavior data importance index and its initial weight, and a dynamic sampling rate is set according to the importance score. Adaptive sampling is performed through the dynamic sampling rate to obtain user behavior data.

3. The e-commerce intelligent recommendation system based on user behavior big data according to claim 2 is characterized by: When constructing a user behavior sequence, the data pre-processing unit (2) handles the problem of interruption of the behavior sequence caused by long-term inactivity as follows: Set a time interval threshold. When the time interval between two adjacent behavior nodes in a user behavior sequence is greater than the time interval threshold, it is determined that the sequence is interrupted. For interrupted sequences, a sequence filling strategy is adopted. That is, the behavioral characteristics before the interruption are analyzed. If the previous behavior is concentrated on browsing products in a certain category, several similar products that the user has browsed in the past are selected and inserted into the interruption according to the browsing probability to simulate the user's subsequent behavior.

4. The e-commerce intelligent recommendation system based on user behavior big data according to claim 3 is characterized by: The data preprocessing unit (2) adopts a knowledge graph-based encoding method when encoding product features and behavior features. The specific steps are as follows: Build an e-commerce knowledge graph. Nodes include products, brands, categories, and attributes, and edges represent the relationships between them. For product features, find the corresponding category, brand, and key attribute nodes based on the knowledge graph. Concatenate the embedding vectors of these nodes to form the encoding vector of the product features. For behavioral features, they are mapped into corresponding behavior vectors through one-hot encoding, and then fused with the product feature encoding vector of the behavior node to obtain the final encoding vector containing product and behavior information.

5. The e-commerce intelligent recommendation system based on user behavior big data according to claim 4 is characterized by: The model construction training unit (3) includes an extraction learning module (31). The multi-layer long short-term memory network unit in the extraction learning module (31) introduces a gating mechanism when performing feature extraction and learning. The specific operation is as follows: Based on the input gate, forget gate, and output gate of the long short-term memory network, an adaptive adjustment gate is added. For each time step in the input sequence, the input vector and the hidden state of the previous moment are set, and the outputs of the input gate, forget gate, and output gate are calculated through linear transformation and sigmoid function. Among them, the output gate is used to control how much new information can enter the cell state, the forget gate is used to control how much information of the cell state at the previous moment needs to be forgotten, and the output gate controls how much information of the cell state is output to the current hidden state. Finally, the output of the adaptive adjustment gate is calculated through linear transformation and hyperbolic tangent function. The adaptive adjustment gate dynamically adjusts feature extraction according to the current input and the previous state.

6. The e-commerce intelligent recommendation system based on user behavior big data according to claim 1 is characterized by: The model construction training unit (3) includes a reinforcement learning module (33). When the reinforcement learning module (33) adopts a deep Q network as a reinforcement learning algorithm, the specific method of defining states, actions and rewards is as follows: The state consists of the current user's interest representation vector, the user's current e-commerce page information, and the user's historical behavior statistical characteristics; Select a certain number of products from the product library as recommended candidate actions, with each action corresponding to a set of candidate products; If the user makes a purchase within a certain period of time after seeing the recommended product, a positive reward will be given. If the user clicks on the recommended product to enter the details page, a half positive reward will be given. If the user does not respond to the recommended product, a negative reward will be given.

Images