Commodity recommendation method based on consumer behaviors

Abstract

The invention relates to the technical field of recommendation methods, in particular to a consumer behavior-based commodity recommendation method, which comprises the following steps of: extracting behavior sequence data including clicking, browsing, searching and purchasing from a user interaction behavior log in an e-commerce platform, and carrying out sequence data slicing on the behavior sequence data through a sequence data slicing technology; dividing the long-time behavior sequence of the user into a plurality of behavior segments with a long-term dependency relationship; generating a behavior context feature vector; based on the long-term dependency feature vector and the behavior context feature vector, constructing a context-based intention speculation model, speculating the long-term shopping intention of the user, and generating an intention vector of the user behavior; and through a multi-target optimization algorithm, generating a recommendation strategy meeting various targets, and applying the generated recommendation strategy to real-time recommendation. According to the method and the device, the personalized degree of recommendation is improved, and diversified requirements of the user can be better met, so that the satisfaction degree of the user is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of recommendation methods, and in particular to a commodity recommendation method based on consumer behavior. Background Art

[0002] With the development of e-commerce platforms, data-driven recommendation systems based on consumer behavior have become an important tool for improving user experience and increasing sales. Traditional recommendation systems usually rely on collaborative filtering, content-based recommendation, and hybrid recommendation algorithms, which can improve the accuracy and personalization of recommendations to a certain extent. However, these traditional technologies still have significant shortcomings when faced with complex user behavior patterns and diverse user needs.

[0003] First, existing technologies mainly focus on users’ short-term behaviors and often ignore the long-term dependencies in user behaviors. This makes it difficult for recommendation systems to capture deep patterns in user behaviors and accurately predict users’ long-term shopping intentions. In addition, when traditional recommendation systems process multimodal data (such as text, images, videos, user reviews, etc.), they often process these data sources separately and lack deep fusion of multimodal information, which limits the recommendation system’s comprehensive understanding of user intentions.

[0004] Secondly, existing recommendation strategies usually focus on optimizing a single objective, such as recommendation accuracy. However, in actual applications, recommendation systems need to balance multiple objectives (such as user satisfaction, system response speed, etc.). This makes it difficult for traditional methods to cope with complex recommendation needs, especially in scenarios with high real-time requirements. It is difficult for recommendation strategies to quickly adapt to changes in user behavior, resulting in poor recommendation results. Summary of the invention

[0005] Based on the above objectives, the present invention provides a commodity recommendation method based on consumer behavior.

[0006] The commodity recommendation method based on consumer behavior includes the following steps: S1, behavior sequence data collection: extract behavior sequence data including click, browse, search, and purchase from the user interaction behavior logs in the e-commerce platform, and divide the user's long-term behavior sequence into multiple behavior segments with long-term dependencies through sequence data slicing technology; S2, multimodal behavior context fusion: for the user's behavior sequence data, capture the multimodal data in the user interaction, the multimodal data includes text description, product pictures, video preview, user evaluation, and use the fusion algorithm based on the multimodal attention mechanism to deeply fuse the multimodal data to generate a behavior context feature vector containing semantic understanding, sentiment analysis and visual features. ; S3, long-term intention inference: Use the long short-term memory network LSTM to model behavior fragments with long-term dependencies and capture the long-term dependency feature vector of user behavior. , based on the long-term dependency feature vector and the behavior context feature vector , build a context-based intention inference model to infer users' long-term shopping intentions and generate intention vectors of user behaviors ; S4, Multi-objective optimization and recommendation strategy generation: Intent vector based on user behavior ,Through a multi-objective optimization algorithm, a recommendation strategy that meets multiple objectives is generated, and the generated recommendation strategy is applied to real-time recommendation.

[0007] Optionally, the behavior sequence data collection in S1 includes extracting user interaction behavior logs, obtaining interaction behavior logs generated by users on the platform from the backend database of the e-commerce platform through log analysis tools, wherein the logs include the user's click behavior, browsing history, search queries, and purchase transactions, arranging the behavior sequence data in chronological order, and archiving them according to the user's unique identifier to form a complete behavior sequence of the user; The sequence data slicing technology is based on the time interval between behaviors and the continuity of behavior types. Through a dynamic window adjustment algorithm, a long-term behavior sequence is divided into multiple short-term behavior segments, while ensuring that there is a significant long-term dependency between the behavior segments.

[0008] Optionally, the dynamic window adjustment algorithm divides the long-term behavior sequence into multiple short-term behavior segments, specifically including: The time interval between actions is expressed as: ,in, Indicates two adjacent behaviors and the time interval between Dynamic adjustment of window size: adjust the window size according to the time interval between behaviors and the behavior type. The adjustment formula is: ,in, and is the weight factor of the time interval, is the association metric function between behavior types; Behavior segmentation: Using dynamically adjusted windows Divide user behavior sequences into multiple segments ,in For the A behavior fragment, contained in the window All behavioral data within; It also includes association analysis and segment adjustment: after the initial slicing, the behavioral segments are identified and adjusted through association analysis to ensure that the long-term dependencies between segments are fully preserved, and the segmented behavioral segments are analyzed for association and the inter-segment dependencies are calculated. If the correlation exceeds the preset threshold , then the fragment and merge.

[0009] Optionally, the multimodal behavior context fusion in S2 specifically includes.

[0010] S21, feature extraction: Feature extraction of text descriptions: using natural language processing techniques to convert text into feature vectors : ,in, is the text entered by the user, It is a semantic feature vector extracted from text; the BERT model is used to extract text features; Feature extraction of product images: Use convolutional neural network (CNN) to process product images, extract visual features, and obtain feature vectors : ,in, It is a product picture. is the visual feature vector extracted from the image; Feature extraction of video preview: Use the time series convolution network (TCN) to analyze the video preview, extract the time series features, and obtain the feature vector : ,in, is the video data, is the time series feature vector extracted from the video; Feature extraction of user reviews: Perform sentiment analysis on user review text to obtain sentiment feature vectors : ,in, Is the user review text, is the sentiment feature vector extracted from the reviews; S22, multimodal attention mechanism fusion: The feature vectors extracted from different modal data are fused, and the multimodal attention mechanism is used to assign weights to different modalities: Attention weight calculation: feature vector for each modality Calculating attention weights : ,in, is modal The feature vectors include , , as well as , is the weight matrix, representing the modality The importance of integration, is the total number of modes; Multimodal feature fusion: The feature vectors of each modality are weighted and summed according to the attention weight to generate a fusion vector ,in, is the fused feature vector, which contains the weighted information of all modal data; S23, Generate context feature vector of user behavior : Fusion Vector Contains feature information of all modalities, but in order to further improve the accuracy of recommendation, the fusion vector needs to be processed to generate the context feature vector of user behavior , the fusion vector Input into a fully connected neural network layer and processed by the ReLU nonlinear activation function to generate the final context feature vector .

[0011] Optionally, the LSTM network in S3 is based on the behavior fragment , using the long short-term memory network LSTM to analyze the behavior fragments Modeling is performed to capture the long-term dependency characteristics of user behavior, expressed as: ,in, Represents the long-term dependency feature vector extracted from the behavior fragment.

[0012] Optionally, S3 further includes converting the long-term dependency feature vector and behavior context feature vector Fusion combines the user's current behavior context and the long-term dependency of historical behavior to construct a comprehensive feature vector: ,in, It is a fused comprehensive feature vector, including the long-term dependency of user behaviors and the current behavior context.

[0013] Optionally, the context-based intention inference model construction includes using the fused comprehensive feature vector As input, infer the user's long-term shopping intention and generate the intention vector of the user's behavior : ,in, It is the intention vector of user behavior, which represents the user's shopping intention in the long-term behavior sequence and the current situation. The generated user intention vector It integrates the user's long-term behavioral dependencies and current behavioral context features.

[0014] Optionally, the intention prediction model is based on a feedforward neural network , the model structure includes: Input layer: The input is a comprehensive feature vector , the shape is , represents the dimension of the fusion vector; Hidden layer: one or more fully connected layers, each with a weight matrix and the bias vector , and activated by the ReLU nonlinear activation function: ,in is the output of the previous layer, is the weight matrix of the current layer, is the bias vector; The output layer is a fully connected layer used to generate the user’s intent vector : ,in and is the weight matrix and bias of the output layer, is the output of the last hidden layer.

[0015] Optionally, the multi-objective optimization and recommendation strategy generation in S4 specifically includes: S41, Objective Function Definition: Intent Vector Based on User Behavior , construct the objective function of multi-objective optimization; S42, use the multi-objective optimization algorithm based on Pareto optimal solution to optimize multiple objective functions, generate a Pareto optimal solution set, each optimal solution in the optimal solution set corresponds to a recommended strategy, achieve a balance between different objectives, and find a set of solution sets in the objective space, so that on this solution set, the improvement of any objective will lead to the deterioration of at least one other objective, and optimize the objective function: ,in, It is A Pareto optimal recommendation strategy; S43, Recommended strategy selection and application: Select an optimal strategy in the Pareto optimal solution set According to the current user preferences and system resource conditions, the recommendation accuracy, user satisfaction and system response speed are comprehensively considered, and the recommendation strategy with the best balance point is selected. Applied to real-time recommendation systems, recommendation lists are generated in real time and pushed to users.

[0016] Optionally, the objective function includes: Recommended accuracy : Maximize the proportion of items in the recommendation list that users are actually interested in; Customer satisfaction : Optimize the matching degree between recommendation results and user preferences, and quantify based on user feedback (click rate, dwell time, purchase behavior); System response speed : Minimize the time for recommendation strategy generation and response.

[0017] Beneficial effects of the present invention: The present invention, by introducing sequence data slicing technology, divides the user's long-term behavior sequence into multiple behavior segments with long-term dependencies. The behavior slicing method combines dynamic window adjustment and association analysis, and can effectively capture the deep-level patterns in user behavior. By analyzing the long-term and short-term dependencies of these segments, the LSTM model is used to further extract the potential relationship between behaviors, ensuring that the recommendation system can accurately identify the user's long-term shopping intentions. Therefore, while maintaining a comprehensive analysis of the long-term and short-term dependencies of user behaviors, the method greatly improves the accuracy and relevance of the recommendation results.

[0018] The present invention uses a multimodal attention mechanism to deeply fuse users' multimodal interaction data (such as text descriptions, product images, video previews, and user reviews) on the e-commerce platform to generate a behavioral context feature vector that includes semantic understanding, sentiment analysis, and visual features. By combining these multimodal features with the long-term dependency features of user behavior fragments, a context-based intention inference model is constructed to further accurately infer users' long-term shopping intentions. This fusion method significantly improves the personalization of recommendations and can better meet users' diverse needs, thereby improving user satisfaction.

[0019] The present invention introduces a multi-objective optimization algorithm based on Pareto optimal solution to simultaneously optimize multiple objectives such as recommendation accuracy, user satisfaction and system response speed. By utilizing these algorithms in the generation of recommendation strategies, an effective balance between multiple objectives can be achieved, and the optimal recommendation strategy that meets the needs of different users and systems can be generated, so that the recommendation strategy can quickly adapt to changing user behaviors and system states, thereby improving the overall performance and user experience of the recommendation system. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings in the following description are only for the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0021] Figure 1 A schematic diagram of a recommended method flow chart of an embodiment of the present invention; Figure 2 It is a schematic diagram of an intention inference model according to an embodiment of the present invention. DETAILED DESCRIPTION

[0022] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. At the same time, it is explained here that in order to make the embodiments more detailed, the following embodiments are the best and preferred embodiments, and those skilled in the art may also adopt other alternatives to implement some known technologies; and the accompanying drawings are only for more specific description of the embodiments, and are not intended to specifically limit the present invention.

[0023] It should be noted that the references to "one embodiment", "embodiment", "exemplary embodiments", "some embodiments" and the like in the specification indicate that the embodiments described may include specific features, structures or characteristics, but not every embodiment may include the specific features, structures or characteristics. In addition, when a specific feature, structure or characteristic is described in conjunction with an embodiment, it should be within the knowledge of a person skilled in the art to implement such feature, structure or characteristic in conjunction with other embodiments (whether or not explicitly described).

[0024] In general, a term can be understood, at least in part, from its use in context. For example, depending, at least in part, on the context, the term "one or more" as used herein can be used to describe any feature, structure, or characteristic in the singular sense, or can be used to describe a combination of features, structures, or characteristics in the plural sense. Additionally, the term "based on" can be understood as not necessarily intended to convey an exclusive set of factors, but can instead, depending, at least in part, on the context, allow for the presence of other factors that are not necessarily explicitly described.

[0025] like Figure 1-Figure 2 As shown, the commodity recommendation method based on consumer behavior includes the following steps: S1, behavior sequence data collection: extract behavior sequence data including click, browse, search, and purchase from the user interaction behavior logs in the e-commerce platform, and divide the user's long-term behavior sequence into multiple behavior segments with long-term dependencies through sequence data slicing technology, so as to ensure that user behaviors at different time points can be analyzed by correlation. Long-term dependencies are based on dynamic window adjustment algorithm and correlation analysis to identify potential dependencies between user behaviors; S2, multimodal behavior context fusion: for the user's behavior sequence data, capture the multimodal data in the user interaction, the multimodal data includes text description, product pictures, video preview, user evaluation, and use the fusion algorithm based on the multimodal attention mechanism to deeply fuse the multimodal data to generate a behavior context feature vector containing semantic understanding, sentiment analysis and visual features. , ensuring that rich contextual information of users is considered in behavioral analysis, enhancing the recommendation system’s understanding of user intent; S3, long-term intention inference: Use the long short-term memory network LSTM to model behavior fragments with long-term dependencies and capture the long-term dependency feature vector of user behavior. , based on the long-term dependency feature vector and the behavior context feature vector , build a context-based intention inference model to infer users' long-term shopping intentions and generate intention vectors of user behaviors ; S4, Multi-objective optimization and recommendation strategy generation: Intent vector based on user behavior ,Through a multi-objective optimization algorithm, a recommendation strategy that meets multiple objectives (such as recommendation accuracy, user satisfaction, system response speed, etc.) is generated, and the generated recommendation strategy is applied to real-time recommendation.

[0026] The behavior sequence data collection in S1 includes the extraction of user interaction behavior logs. The interaction behavior logs generated by users on the platform are obtained from the backend database of the e-commerce platform through log analysis tools. The logs include the user's click behavior, browsing history, search queries, and purchase transactions. The behavior sequence data is arranged in chronological order and archived according to the user's unique identifier to form a complete behavior sequence of the user; Sequence data slicing technology slices data based on the time interval between behaviors and the continuity of behavior types. Through a dynamic window adjustment algorithm, long-term behavior sequences are divided into multiple short-term behavior segments, while ensuring that there are significant long-term dependencies between the behavior segments. The slicing process also takes into account the correlation between different behavior segments on the timeline, so as to effectively identify potential dependencies between user behaviors in subsequent analysis.

[0027] Dynamic window adjustment is the basis for ensuring long-term dependencies between behavior segments. Specifically, the dynamic window adjustment algorithm dynamically determines the length of each slice based on the time interval and behavior type in the behavior sequence.

[0028] Time interval analysis: The size of the window depends on the time interval Δtk between behaviors. When the time interval between behaviors is short (such as when a user browses multiple products in a short period of time), these behaviors may have a strong correlation, so they are divided into the same segment.

[0029] Behavior type association: There may also be dependencies between different types of behaviors, such as a user searching for a product and then immediately purchasing it. By identifying common patterns or association rules between different behavior types (e.g., clicks are often followed by purchases), we can ensure that these behaviors are classified in the same segment.

[0030] Through the above two points, the dynamic window adjustment algorithm can automatically adapt to the time and type changes between behaviors, thereby dividing behaviors with potential dependencies into the same segment.

[0031] The slicing process not only depends on the initial time interval and behavior type, but also performs feedback optimization through the results of association analysis: after initial slicing, the potential dependencies in the behavior fragments are identified through association analysis. If the correlation between certain behavior fragments is high (high support and confidence), it means that they should be classified as the same fragment. In this case, the slicing algorithm will backtrack and adjust the slicing strategy to merge these behaviors.

[0032] Balancing short- and long-term dependencies: When slicing, it is also necessary to balance short-term behavior patterns and long-term dependencies. To this end, the algorithm performs multi-scale analysis based on the time span to ensure that even if the behavior segments span a large span, the dependencies between them can be identified and retained. For example, some user behaviors may recur after a period of time. This long-term dependency can be retained by adjusting the window size and interval when slicing.

[0033] Representation of user behavior sequence: ,in, Indicates user The behavior sequence, Expressing behavior The timestamp of occurrence, Represents a specific type of behavior, such as click, browse, search, or purchase; by timestamp right Sort all behaviors in the data to ensure that the user's behavior sequence is arranged in chronological order.

[0034] The dynamic window adjustment algorithm divides the long-term behavior sequence into multiple short-term behavior segments, including: The time interval between actions is expressed as: ,in, Indicates two adjacent behaviors and the time interval between Dynamic adjustment of window size: Indicates the time length of the slice window, defined For the The time window of the slices, The initial representation is: , Indicates the number of behaviors in the window. The window size is adjusted according to the time interval between behaviors and the behavior type. The adjustment formula is: ,in, and is the weight factor of the time interval, is the association metric function between behavior types, expressed as: ; Here, the support and confidence are calculated as follows: Support Confidence in, is the weight factor of the behavior type relevance, It is the preset maximum window size, which is used to limit the maximum span of the window.

[0035] Behavior segmentation: Using dynamically adjusted windows Divide user behavior sequences into multiple segments ,in For the A behavior fragment, contained in the window All behavioral data within; It also includes association analysis and segment adjustment: after the initial slicing, the behavioral segments are identified and adjusted through association analysis to ensure that the long-term dependencies between segments are fully preserved, and the segmented behavioral segments are analyzed for association and the inter-segment dependencies are calculated. Degree of association: ,in It's a fragment and The correlation between Indicates that in the user behavior sequence, behavior Behavior that occurs immediately after For example, It can be the behavior of a user clicking on a product, is the user's subsequent purchase of the product. If the correlation exceeds the preset threshold , then the fragment and Merge, expressed as: , Indicates user No. A behavioral fragment.

[0036] The multimodal behavior context fusion in S2 specifically includes: During the user interaction process, multimodal data of the user's interaction with the platform is captured and recorded in real time, including: Text description: text data entered by users when searching, commenting, or filling out forms; Product images: images of products that users click or browse; Video preview: product promotion video or usage tutorial watched by users; User reviews: Ratings and comments from users on products after purchase.

[0037] S21, feature extraction: Feature extraction of text descriptions: using natural language processing techniques to convert text into feature vectors : ,in, is the text entered by the user, It is a semantic feature vector extracted from text; use the BERT model to extract text features and select the BERT language model. The BERT language model is a pre-trained language model based on the bidirectional Transformer, which can extract deep semantic features from text. The model input is the text T input by the user. The model structure includes the input layer: the text T is converted into a word vector sequence through the word embedding layer, and the BERT encoder: the BERT model consists of a multi-layer Transformer encoder, each layer contains a multi-head self-attention mechanism and a feedforward neural network. The text is contextually understood through these layers to obtain the contextual representation of each word. The output layer uses the output vector corresponding to the BERT [CLS] tag as the semantic feature vector of the entire sentence. ; Feature extraction of product images: Use convolutional neural network (CNN) to process product images, extract visual features, and obtain feature vectors : ,in, It is a product picture. It is a visual feature vector extracted from the image. The product image I is input, and the local features of the image are extracted through multi-layer convolution operations. Each convolution layer is followed by a ReLU activation function and a pooling layer for nonlinear transformation and downsampling of features. After the feature map output by the convolution layer is flattened, it is input into the fully connected layer to extract global visual features. Finally, the output layer outputs a fixed-length visual feature vector ; Feature extraction of video preview: Use the time series convolution network (TCN) to analyze the video preview, extract the time series features, and obtain the feature vector : ,in, is the video data, It is the time series feature vector extracted from the video; input video data V, extract features for each frame of the video, obtain a frame-level feature vector sequence, input the frame-level feature sequence into the TCN network, capture the time dependency between video frames, extract the time series features, and output the final hidden state of the TCN as the video feature vector ; Feature extraction of user reviews: Perform sentiment analysis on user review text to obtain sentiment feature vectors : ,in, Is the user review text, It is the sentiment feature vector extracted from the evaluation; the sentiment analysis model is used to extract the sentiment features in the user evaluation. The sentiment analysis model is trained based on the BERT model. The model inputs the user evaluation text R, uses the BERT model to encode the evaluation text, extracts the sentiment features in the text, and uses the classifier to identify the sentiment tendency of the text (such as positive, negative, neutral) based on the BERT output. The output layer extracts the sentiment-related feature vector ; S22, multimodal attention mechanism fusion: The feature vectors extracted from different modal data are fused, and the multimodal attention mechanism is used to assign weights to different modalities: Attention weight calculation: feature vector for each modality Calculating attention weights : ,in, is modal The feature vectors include , , as well as , is the weight matrix, representing the modality The importance of integration, is the total number of modalities (here 4, representing text, image, video, and review); Multimodal feature fusion: The feature vectors of each modality are weighted and summed according to the attention weight to generate a fusion vector ,in, is the fused feature vector, which contains the weighted information of all modal data; S23, Generate context feature vector of user behavior : Fusion Vector Contains feature information of all modalities, but in order to further improve the accuracy of recommendation, the fusion vector needs to be processed to generate the context feature vector of user behavior , the fusion vector Input into a fully connected neural network layer and processed by the ReLU nonlinear activation function to generate the final context feature vector : ,in, is the weight matrix of the fully connected layer, is the bias vector, is the ReLU nonlinear activation function, .

[0038] Long Short-Term Memory Network (LSTM) in S3 based on behavior fragments , using the long short-term memory network LSTM to analyze the behavior fragments Modeling is performed to capture the long-term dependency characteristics of user behavior, expressed as: ,in, Represents the long-term dependency feature vector extracted from the behavior fragment.

[0039] The LSTM model structure is as follows: Input layer: The input is a sequence of multiple behavior fragments , each fragment is a vector representing the features extracted from this fragment, with a shape of ,in is the number of fragments, is the dimension of the fragment feature vector. LSTM unit, including the following: Forget Gate: , the Forget Gate Controls how much state information from the previous moment is discarded. is the hidden state of the previous LSTM unit, is the current input segment feature vector, and are the weight matrix and bias vector; Input Gate: , input gate decides how much new information to add to the current cell's state, and are the corresponding weights and biases; Candidate status update: , candidate status It is the temporary state information updated by inputting the current segment features and the hidden state of the previous moment; Update cell status: ; Cell state The weighted sum of the previous state and the current state is adjusted by the forget gate and the input gate; Output Gate: , output gate Determine the hidden state output at the current moment; Hide status updates: , hidden state is the output of the current LSTM unit and is passed to the next time step.

[0040] Output layer: final hidden state is used as a feature vector of long-term dependencies between behavioral segments: .

[0041] S3 also includes the long-term dependency feature vector and behavior context feature vector Fusion combines the user's current behavior context and the long-term dependency of historical behavior to construct a comprehensive feature vector: ,in, It is a fused comprehensive feature vector, including the long-term dependencies of user behaviors and the current behavior context; Concatenate means connecting two or more vectors by dimension to form a new, larger vector, which is used to combine features from different sources so as to be input into subsequent models for further processing.

[0042] The context-based intention inference model construction includes using the fused comprehensive feature vector As input, infer the user's long-term shopping intention and generate the intention vector of the user's behavior : ,in, It is the intention vector of user behavior, which represents the user's shopping intention in the long-term behavior sequence and the current situation. The generated user intention vector It integrates the user's long-term behavioral dependencies and current behavioral context features as the key input for subsequent recommendation strategy generation, providing the system with more accurate personalized recommendations.

[0043] Intent prediction model based on feedforward neural network , the model structure includes: Input layer: The input is a comprehensive feature vector , the shape is , represents the dimension of the fusion vector; Hidden layer: one or more fully connected layers, each with a weight matrix and the bias vector , and activated by the ReLU nonlinear activation function: ,in is the output of the previous layer, is the weight matrix of the current layer, is the bias vector; The output layer is a fully connected layer used to generate the user’s intent vector : ,in and is the weight matrix and bias of the output layer, is the output of the last hidden layer.

[0044] The multi-objective optimization and recommendation strategy generation in S4 specifically include: S41, Objective Function Definition: Intent Vector Based on User Behavior , construct the objective function of multi-objective optimization; S42, use the multi-objective optimization algorithm based on Pareto optimal solution to optimize multiple objective functions, generate a Pareto optimal solution set, each optimal solution in the optimal solution set corresponds to a recommended strategy, achieve a balance between different objectives, and find a set of solution sets in the objective space, so that on this solution set, the improvement of any objective will lead to the deterioration of at least one other objective, and optimize the objective function: ,in, It is A Pareto optimal recommendation strategy; S43, Recommended strategy selection and application: Select an optimal strategy in the Pareto optimal solution set ,According to the current user preferences and system resources, taking into account the recommendation accuracy, user satisfaction and system response speed factors, the recommendation strategy with the best balance point is selected. The strategy selection formula is: ,in It is the weight parameter of different objectives, which is dynamically adjusted according to actual needs and resource conditions. Argmax is a mathematical operator, which represents the input value that makes the objective function take the maximum value in a certain function. It is used here to select the best recommendation strategy. Applied to real-time recommendation systems, recommendation lists are generated in real time and pushed to users.

[0045] The objective functions include: Recommended accuracy :Maximize the proportion of items in the recommendation list that users are actually interested in: Customer satisfaction : Optimize the matching degree between recommendation results and user preferences, and quantify based on user feedback (click rate, dwell time, purchase behavior): ,in, It is Recommended products, is the number of items in the recommendation list; System response speed : Minimize the time of recommendation strategy generation and response: .

[0046] The multi-objective optimization algorithm based on Pareto optimal solution optimizes multiple objective functions and generates a Pareto optimal solution set. The specific process is as follows, using NSGA-II optimization process: Initialize the population: Generate N initial solutions, each of which is a recommended strategy.

[0047] Non-dominated sorting: non-dominated sorting of the population according to multiple objective functions to generate multiple frontiers; Selection, crossover and mutation: select individuals from the non-dominated frontier for crossover and mutation to generate the next generation population; Iteration: Repeat the non-dominated sorting and crowding distance calculation until the termination condition is reached.

[0048] Generate Pareto optimal solution set: The final solution set is the Pareto optimal solution set, each solution of which represents a different equilibrium point.

[0049] 1. Non-dominated sorting (NSGA-II characteristics): Perform non-dominated sorting on the individuals in the population. Non-dominated sorting divides the individuals into multiple levels (frontiers), where the first frontier contains non-dominated individuals (optimal solutions).

[0050] In a non-dominated sorting, an individual Ra dominates an individual Rb if and only if Ra is not inferior to Rb in all objectives and is superior to Rb in at least one objective.

[0051] 2. Crowding distance calculation (NSGA-II feature): The crowding distance of each individual is calculated to maintain the diversity of the population. The crowding distance is determined by calculating the distance between each individual and its adjacent individuals. The larger the distance, the sparser the solution in the area where the individual is located, and thus the greater the probability of being retained.

[0052] The present invention covers any substitution, modification, equivalent method and scheme made on the essence and scope of the present invention. In order to make the public have a thorough understanding of the present invention, specific details are described in detail in the following preferred embodiments of the present invention, but those skilled in the art can fully understand the present invention without the description of these details. In addition, in order to avoid unnecessary confusion about the essence of the present invention, well-known methods, processes, procedures, components and circuits are not described in detail.

[0053] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A commodity recommendation method based on consumer behavior, characterized in that: The following steps are involved: S1, behavior sequence data collection: extract behavior sequence data including click, browse, search, and purchase from the user interaction behavior logs in the e-commerce platform, and divide the user's long-term behavior sequence into multiple behavior segments with long-term dependencies through sequence data slicing technology; S2, multimodal behavior context fusion: for the user's behavior sequence data, capture the multimodal data in the user interaction, the multimodal data includes text description, product pictures, video preview, user evaluation, and use the fusion algorithm based on the multimodal attention mechanism to deeply fuse the multimodal data to generate a behavior context feature vector containing semantic understanding, sentiment analysis and visual features; S3, long-term intention inference: Use the long short-term memory network (LSTM) to model behavior fragments with long-term dependencies, capture the long-term dependency feature vector of user behavior, and build a context-based intention inference model based on the long-term dependency feature vector and the behavior context feature vector to infer the user's long-term shopping intention and generate the user's behavior intention vector. S4, multi-objective optimization and recommendation strategy generation: Based on the intention vector of user behavior, a multi-objective optimization algorithm is used to generate a recommendation strategy that meets multiple objectives, and the generated recommendation strategy is applied to real-time recommendation.

2. The method for recommending products based on consumer behavior according to claim 1, characterized in that: The behavior sequence data collection in S1 includes the extraction of user interaction behavior logs, and the interaction behavior logs generated by users on the platform are obtained from the backend database of the e-commerce platform through log analysis tools. The logs include the user's click behavior, browsing history, search queries, and purchase transactions. The behavior sequence data are arranged in chronological order and archived according to the user's unique identifier to form a complete behavior sequence of the user; The sequence data slicing technology is based on the time interval between behaviors and the continuity of behavior types. Through a dynamic window adjustment algorithm, a long-term behavior sequence is divided into multiple short-term behavior segments, while ensuring that there are long-term dependencies between the behavior segments.

3. The method for recommending products based on consumer behavior according to claim 2, characterized in that: The dynamic window adjustment algorithm divides a long-term behavior sequence into multiple short-term behavior segments, specifically including: The time interval between actions is expressed as: ,in, Indicates two adjacent behaviors and the time interval between Dynamic adjustment of window size: adjust the window size according to the time interval between behaviors and the behavior type; Behavior segmentation: Use dynamically adjusted windows to divide the user behavior sequence into multiple segments, where each segment contains all the behavior data within the window; It also includes association analysis and fragment adjustment: after preliminary slicing, the behavioral fragments are identified and adjusted through association analysis to ensure that the long-term dependencies between fragments are fully preserved. The sliced ​​behavioral fragments are subjected to association analysis to calculate the correlation between fragments. If the correlation exceeds the preset threshold, the two adjacent fragments are merged.

4. The method for recommending products based on consumer behavior according to claim 3, characterized in that: The multimodal behavior context fusion in S2 specifically includes: S21, feature extraction: Feature extraction of text description: Use natural language processing technology to convert text into text feature vectors, and use the BERT model to extract text features; Feature extraction of product images: Use convolutional neural network (CNN) to process product images, extract visual features, and obtain visual feature vectors; Feature extraction of video preview: Use the temporal convolutional network (TCN) to analyze the video preview, extract the time series features, and obtain the time series feature vector; Feature extraction of user reviews: Perform sentiment analysis on user review text to obtain sentiment feature vectors; S22, multimodal attention mechanism fusion: The feature vectors extracted from different modal data are fused, and the multimodal attention mechanism is used to assign weights to different modalities: Attention weight calculation: Combine the weight matrix with the feature vector of each modality Calculating attention weights , is modal The feature vectors include , , as well as ; Multimodal feature fusion: The feature vectors of each modality are weighted and summed according to the attention weights to generate a fusion vector, which contains the weighted information of all modal data and the feature information of all modalities; S23, generating a context feature vector of the user behavior: processing the fusion vector to generate a context feature vector of the user behavior, inputting the fusion vector into a fully connected neural network layer, processing it through a ReLU nonlinear activation function, and generating a final context feature vector.

5. The method for recommending products based on consumer behavior according to claim 4, characterized in that: The LSTM in S3 is based on behavior fragments, and the LSTM is used to model the behavior fragments to capture the long-term dependency characteristics of user behavior.

6. The method for recommending products based on consumer behavior according to claim 5, characterized in that: The S3 also includes fusing the long-term dependency feature vector with the behavior context feature vector, combining the user's current behavior context and the long-term dependency of historical behavior to construct a comprehensive feature vector, wherein the comprehensive feature vector includes the long-term dependency of the user's behavior and the current behavior context.

7. The method for recommending products based on consumer behavior according to claim 6, characterized in that: The construction of the context-based intention inference model includes using the fused comprehensive feature vector as input to infer the user's long-term shopping intention and generate a user behavior intention vector. The user behavior intention vector represents the user's shopping intention in the long-term behavior sequence and the current situation. The generated user intention vector integrates the user's long-term behavior dependency and current behavior context characteristics.

8. The method for recommending products based on consumer behavior according to claim 7, characterized in that: The intention prediction model is based on a feed-forward neural network , the model structure includes: Input layer: The input is the comprehensive feature vector, the shape is , represents the dimension of the fusion vector; Hidden layer: one or more fully connected layers, each layer has a weight matrix and bias vector, and is activated by the ReLU nonlinear activation function; The output layer is a fully connected layer, which is used to generate the user's intention vector. The product of the output of the last hidden layer and the weight matrix of the output layer plus the bias of the output layer is the user's intention vector.

9. The method for recommending products based on consumer behavior according to claim 1, characterized in that: The multi-objective optimization and recommendation strategy generation in S4 specifically include: S41, Objective function definition: Based on the intention vector of user behavior, construct the objective function of multi-objective optimization; S42, use the multi-objective optimization algorithm based on Pareto optimal solution to optimize multiple objective functions, generate a Pareto optimal solution set, each optimal solution in the optimal solution set corresponds to a recommended strategy, achieve a balance between different objectives, and find a set of solution sets in the objective space, so that on this solution set, the improvement of any objective will lead to the deterioration of at least one other objective, and optimize the objective function: ,in, It is A Pareto optimal recommendation strategy; S43, Recommended strategy selection and application: Select an optimal strategy in the Pareto optimal solution set According to the current user preferences and system resource conditions, the recommendation accuracy, user satisfaction and system response speed are comprehensively considered, and the recommendation strategy with the best balance point is selected. Applied to real-time recommendation systems, recommendation lists are generated in real time and pushed to users.

10. The method for recommending products based on consumer behavior according to claim 9, characterized in that: The objective function includes: Recommended accuracy : Maximize the proportion of items in the recommendation list that users are actually interested in; Customer satisfaction : Optimize the matching degree between recommendation results and user preferences, and quantify based on user feedback; System response speed : Minimize the time for recommendation strategy generation and response.

Images