A method and system for recommending products on an e-commerce platform based on user preference analysis

Abstract

The present invention discloses a method and system for recommending goods on an e-commerce platform based on user preference analysis, which relates to the technical field of e-commerce platforms, including: constructing a user multimodal interest feature vector; the multimodal interest feature vector integrates the user's text search history, image browsing preference, audio consumption habits, and video interactive behavior; using a cross-modal attention mechanism, matching the user's multimodal interest feature vector with the multimodal attribute information of goods to generate a candidate product set; applying a dynamic context-aware reinforcement learning model to reorder the candidate product set according to the user's real-time interactive behavior and environmental factors, and outputting the final recommendation result. The present invention realizes comprehensive capture and accurate modeling of user interests by combining multimodal interest feature extraction, dynamic context awareness and reinforcement learning technology.

Description

Technical Field

[0001] The present invention relates to the technical field of e-commerce platforms, and in particular to a method and system for recommending products on an e-commerce platform based on user preference analysis. Background Art

[0002] The product recommendation system of e-commerce platforms plays a vital role in improving user experience and promoting sales. With the continuous advancement of technology, recommendation systems have developed from early rule-based methods to today's complex algorithm systems. At present, the mainstream recommendation technologies mainly include collaborative filtering, content-based recommendation and hybrid recommendation methods. Collaborative filtering technology analyzes the user's historical behavior data to find similar users or similar products, thereby recommending products that may be of interest to the target user. This method can be further divided into user-based collaborative filtering and item-based collaborative filtering. Content-based recommendation focuses on analyzing the attribute characteristics of products and user preferences, and generates recommendation results by calculating the similarity between user preferences and product characteristics. Hybrid recommendation methods combine multiple recommendation strategies to make up for the shortcomings of a single method. In recent years, with the development of deep learning technology, recommendation models based on neural networks have gradually become a research hotspot, such as deep collaborative filtering, attention mechanism and graph neural network. These models can better capture the complex nonlinear relationship between user behavior and product characteristics and improve the accuracy of recommendations. In addition, some studies have begun to focus on the dynamic and context-aware capabilities of recommendation systems, trying to take into account dynamic information such as users' real-time behavior and environmental factors to provide more personalized and timely recommendation services.

[0003] Although existing recommendation technologies have made significant progress, there are still some urgent problems to be solved in the face of increasingly complex e-commerce environments and user needs. First, most existing methods find it difficult to effectively handle the multimodality and dynamic characteristics of user interests. Users' interests on e-commerce platforms are often multifaceted and may involve multiple modalities such as text, images, and videos at the same time, while existing technologies can only handle a single or limited number of modalities, making it difficult to fully capture the characteristics of users' interests. Second, traditional recommendation models are mostly static and cannot respond to users' real-time behaviors and environmental changes in a timely manner. For example, users' interests may change rapidly over time, scenarios, and devices, but existing technologies find it difficult to adjust recommendation strategies in real time to adapt to these changes. Furthermore, existing recommendation systems usually use fixed matching algorithms that lack flexibility and adaptability. As a result, in a complex and changing e-commerce environment, the recommendation effect may gradually decrease over time. In addition, most recommendation methods rely mainly on pre-defined static rules or simple scoring mechanisms when sorting products, making it difficult to fully consider the balance between users' long-term preferences and short-term interests, and unable to effectively use users' feedback information to continuously optimize recommendation strategies. Finally, existing technologies still face challenges in dealing with cold start problems and long-tail product recommendations, making it difficult to provide accurate recommendations for new users or niche products. These problems limit the performance and scope of application of recommendation systems, affecting user experience and the commercial value of the platform. Therefore, there is an urgent need for a new recommendation method that can comprehensively consider user multimodal interests, dynamic context, real-time interactive behavior, and has adaptive learning capabilities to overcome the limitations of existing technologies and provide more accurate, personalized, and real-time product recommendation services. Summary of the invention

[0004] In view of the above-mentioned problems, the present invention is proposed.

[0005] Therefore, the technical problem solved by the present invention is: how to provide an e-commerce platform product recommendation method and system based on user preference analysis, so as to solve the technical problems that the existing e-commerce platform product recommendation system is difficult to effectively handle user multimodal interest characteristics, dynamic context information and real-time interactive behavior, and lacks adaptive learning capabilities, thereby providing more accurate, personalized and real-time product recommendation services.

[0006] To solve the above technical problems, the present invention provides the following technical solutions: a method for recommending products on an e-commerce platform based on user preference analysis, comprising: constructing a user multimodal interest feature vector; the multimodal interest feature vector integrates the user's text search history, image browsing preferences, audio consumption habits, and video interaction behavior; using a cross-modal attention mechanism to match the user's multimodal interest feature vector with the multimodal attribute information of the product to generate a set of candidate products; applying a dynamic context-aware reinforcement learning model to reorder the initial set of candidate products according to the user's real-time interaction behavior and environmental factors, and outputting the final recommendation result.

[0007] As a preferred solution of the e-commerce platform product recommendation method based on user preference analysis described in the present invention, wherein: constructing a user multimodal interest feature vector includes the following steps: extracting user historical browsing records, purchase records and search records from the e-commerce platform database; performing data cleaning on the user historical browsing records, purchase records and search records, removing duplicate data and invalid data, and classifying the cleaned user historical browsing records, purchase records and search records according to product category, brand, and price range; based on the classified historical browsing records, purchase records and search records, calculating the user's preference weights for each product category, brand and price range, and generating a user multimodal interest feature vector.

[0008] As a preferred solution of the e-commerce platform product recommendation method based on user preference analysis described in the present invention, wherein: the cross-modal attention mechanism is used to match the user's multimodal interest feature vector with the multimodal attribute information of the product to generate an initial candidate product set, including the following steps: receiving the user's multimodal interest feature vector; constructing a product feature matrix based on the multimodal attribute information of the product in the e-commerce platform database; at the same time, obtaining the historical behavior sequence of the user's interaction within the historical time; when the user's multimodal interest feature vector, the product feature matrix and the historical behavior sequence are input, the initial attention weight matrix is ​​calculated through the dynamic weight cross-modal attention network; using the temporal attention mechanism based on the The weights in the initial attention weight matrix are adjusted according to the historical behavior sequence to obtain the attention weight matrix; if the value of an element in the attention weight matrix is ​​lower than the sparsification threshold, the element is set to 0 to obtain a sparse attention weight matrix; based on the sparse attention weight matrix, the user-item matching score is calculated: when the user-item matching score exceeds the matching score threshold τ, the corresponding item is added to the initial candidate item set; if the size of the initial candidate item set reaches a preset number or all items have been evaluated, the items in the initial candidate item set are sorted in descending order according to the user-item matching score, and the sorted initial candidate item set is output as the final candidate item set.

[0009] As a preferred solution of the e-commerce platform product recommendation method based on user preference analysis described in the present invention, the user-product matching score is expressed by the following formula:

[0010]

[0011] Among them, S i is the user-item matching score of the i-th item; i is the index of the item; d is the total dimension of the feature, which is equal to the sum of the item category dimension k, the brand dimension l, and the price range dimension m; j is the index of the feature dimension, from 1 to d; A″ ij is the weight value of the jth feature dimension of the ith product in the sparse attention weight matrix A”; g ij is the value of the jth feature dimension of the ith product in the product feature matrix G; z i It is the dot product sum of the attention weight and the product features;

[0012]

[0013] f(A″ ij ,g ij )=ReLU(A″ ij +g ij -|A″ ij -g ij |)

[0014] Among them, f(A″ ij , g ij ) is a nonlinear function used to capture the interaction between attention weights and item features.

[0015] As a preferred solution of the e-commerce platform product recommendation method based on user preference analysis described in the present invention, the dynamic weight cross-modal attention network dynamically captures the relationship between user interests and product features through the attention mechanism, while considering the interaction between product categories, brands, and price ranges.

[0016] As a preferred solution of the e-commerce platform product recommendation method based on user preference analysis described in the present invention, the application of dynamic context-aware reinforcement learning model includes the following steps: constructing a state space, combining the user multimodal interest feature vector, the product feature matrix, the historical behavior sequence H, the current time and the device type to form a state vector. Based on the user multimodal interest feature vector and the user-product matching score, the action space is designed; the state encoder of the multi-head attention mechanism is used to map the state vector to a low-dimensional representation, and the number of attention heads matches the dimension of the user multimodal interest feature vector; using the user-product matching score, a reinforcement learning reward function that comprehensively considers the product matching degree and the user's real-time feedback is designed; based on the low-dimensional representation and the reinforcement learning reward function, the policy network is used to select the optimal action.

[0017] As a preferred solution of the e-commerce platform product recommendation method based on user preference analysis described in the present invention, the candidate product set is reordered according to the optimal action, a sorted product list is generated, and the top K products are selected as the final recommendation results and output to the user.

[0018] To further solve the above technical problems, the present invention provides the following technical solutions: an e-commerce platform product recommendation system based on user preference analysis, comprising: a construction module, used to construct a user multimodal interest feature vector; a generation module, used to use a cross-modal attention mechanism to match the user multimodal interest feature vector with the multimodal attribute information of the product to generate a set of candidate products; an output module, used to apply a dynamic context-aware reinforcement learning model to reorder the candidate product set according to the user's real-time interactive behavior and environmental factors, and output the final recommendation result.

[0019] A computer device includes a memory and a processor, wherein the memory stores a computer program, and is characterized in that when the processor executes the computer program, the steps of the e-commerce platform product recommendation method based on user preference analysis as described above are implemented.

[0020] A computer-readable storage medium having a computer program stored thereon, characterized in that when the computer program is executed by a processor, the steps of the e-commerce platform product recommendation method based on user preference analysis as described above are implemented.

[0021] Beneficial effects of the present invention: The present invention realizes comprehensive capture and accurate modeling of user interests by combining multimodal interest feature extraction, dynamic context perception and reinforcement learning technology. The present invention not only takes into account the user's historical behavior and long-term preferences, but also can respond to the user's immediate needs and environmental changes in real time, thereby performing intelligent re-ranking based on the initial set of candidate products. This innovative combination significantly improves the accuracy, personalization and adaptability of the recommendation system, and effectively solves the problem that traditional recommendation methods perform poorly in dynamic environments. At the same time, by introducing multi-head attention mechanisms and reinforcement learning strategies, the present invention can better capture the complex nonlinear relationship between user interests and product features, further improving the recommendation effect. Ultimately, the present invention can not only provide users with more accurate and personalized product recommendations, but also improve user satisfaction and the commercial value of the platform, which is of great significance to improving the user experience and operational efficiency of the e-commerce platform. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0023] Figure 1 An overall flow chart of a method for recommending products on an e-commerce platform based on user preference analysis provided by one embodiment of the present invention.

[0024] Figure 2 A diagram of a computer device for the e-commerce platform product recommendation method based on user preference analysis of the present invention. DETAILED DESCRIPTION

[0025] In order to make the above-mentioned purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the drawings of the specification. Obviously, the described embodiments are part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in the art without creative work should fall within the scope of protection of the present invention.

[0026] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0027] Example 1, reference Figure 1, which is an embodiment of the present invention, provides an e-commerce platform product recommendation method based on user preference analysis.

[0028] In the prior art,

[0029] To solve the above problems, the present invention provides a method for recommending products on an e-commerce platform based on user preference analysis. Figure 1 As shown, it is an overall flow chart of the e-commerce platform product recommendation method based on user preference analysis provided by the present invention, including:

[0030] S1: Construct a user multimodal interest feature vector, which integrates the user's text search history, image browsing preferences, audio consumption habits, and video interaction behavior.

[0031] S1.1: Extract the user’s historical browsing history, purchase history, and search history from the e-commerce platform database within the past 30 days.

[0032] S1.2: Perform data cleaning on the extracted historical browsing records, purchase records, and search records, remove duplicate data and invalid data, and classify the cleaned historical browsing records, purchase records, and search records according to product category, brand, and price range.

[0033] S1.3: Based on the classified historical browsing records, purchase records and search records, calculate the user's preference weights for each product category, brand and price range, and generate a user interest feature vector.

[0034] Specifically, the calculation formula of the user's multimodal interest feature vector is:

[0035]

[0036] in, is the user's multimodal interest feature vector; is the product category preference vector; Brand preference vector; is the price range preference vector; w1, w2, w3 are the weight coefficients of product category, brand, and price range, respectively, and w1+w2+w3=1.

[0037] Furthermore, the product category preference vector The calculation formula is:

[0038]

[0039] Among them, C i is the preference score of the i-th product category; is the number of times a user browses the i-th product category; is the number of searches by users in the i-th product category; is the number of purchases made by the user in the i-th product category; T v is the total number of user views; T s is the total number of searches by the user; T p is the total number of purchases by the user; α, β, γ are the weight coefficients of browsing, searching, and purchasing behaviors, respectively, and α+β+γ=1.

[0040] Brand Preference Vector and the price range preference vector The calculation method is similar to the product category preference vector

[0041] It should be noted that in order to more accurately reflect the user's recent interest changes, and The time decay factor is introduced:

[0042]

[0043] Where λ is the time decay coefficient; Δt j is the difference between the current time and the time when the jth behavior occurred (in days); count j is the count of the jth behavior (usually 1); n is the total number of behaviors of this category.

[0044] Preferably, step S1 of the present invention significantly improves the accuracy and timeliness of the user interest feature vector by introducing a time decay factor and a dynamic weight adjustment mechanism. Compared with the prior art, this method can more accurately capture the changing trend of users' recent behavioral preferences, and dynamically adjust the weights of various behaviors according to the user's activity level and purchase conversion rate. This improvement enables the generated user interest feature vector to better reflect the user's real-time interests and purchasing tendencies, providing more reliable and personalized input data for subsequent product recommendation algorithms. Especially in the rapidly changing environment of e-commerce platforms, this method can adapt to changes in user interests more quickly, thereby improving the accuracy of recommendations and user satisfaction, which may ultimately lead to increases in click-through rates, conversion rates, and user retention rates.

[0045] S2: Using the cross-modal attention mechanism, the user’s multimodal interest feature vector is matched with the multimodal attribute information of the product to generate an initial set of candidate products.

[0046] S2.1: Receive the user multimodal interest feature vector generated by S1.

[0047] S2.2: Extract the multimodal attribute information of the goods from the e-commerce platform database and construct the product feature matrix G; at the same time, obtain the historical behavior sequence H of the user's interaction in the historical time (the recent N times, which can be set according to actual needs).

[0048] The product feature matrix G is expressed as:

[0049]

[0050] Among them, n is the number of products, k, l, and m are the characteristic dimensions of category, brand, and price range respectively;

[0051] The historical behavior sequence H is expressed as:

[0052] H=[h1,h2,...,h N ];

[0053] Where N is the number of historical behaviors considered, h1,h2,...,h N Indicates h i ,h i The feature vector representing the i-th historical behavior.

[0054] S2.3: When the user's multimodal interest feature vector is input When the product feature matrix G and the historical behavior sequence H are considered, the initial attention weight matrix A is calculated through the dynamic weight cross-modal attention network; then the temporal attention mechanism is used to adjust the weights in the initial attention weight matrix A based on the historical behavior sequence H to obtain the temporally adjusted attention weight matrix A'.

[0055] The dynamic weighted cross-modal attention network is a recommendation system architecture designed specifically for processing multimodal user interest and product feature data in e-commerce platforms. The core idea of ​​the network is to dynamically capture the complex relationship between user interests and product features through the attention mechanism, while considering the interaction between different modalities (such as product category, brand, price range). The main features of the network include:

[0056] Multimodal fusion: The network processes both the user interest feature vector and the product feature matrix, both of which contain information from multiple modalities. Through the attention mechanism, the network can adaptively learn the importance and relevance between different modalities.

[0057] Dynamic weight adjustment: The network introduces a dynamic weight adjustment mechanism, which adaptively adjusts the importance of different feature dimensions through a learnable weight matrix. This enables the network to flexibly adjust the recommendation strategy according to different users and different scenarios.

[0058] Temporal Attention Integration: By introducing the temporal attention mechanism TemporalAttention (H), the network can consider the user's historical behavior sequence, thereby capturing the temporal dynamics of user interests. This design helps to improve the timeliness and personalization of recommendations.

[0059] Adaptive threshold mechanism: The network uses adaptive sparsification thresholds and matching score thresholds, which are dynamically adjusted according to the current attention distribution and historical matching scores. This design enhances the robustness and adaptability of the network, enabling it to better cope with different data distributions and user behavior patterns.

[0060] Modality-specific matching score calculation: When calculating the final user-item matching score, the network maintains the independence of different modalities and uses weights consistent with the user's interest feature vector for weighted summation. This ensures that the recommendation results are highly relevant to the user's multimodal interests.

[0061] Calculation of the initial attention weight matrix A:

[0062]

[0063] Among them, W Q and W K is the learnable weight matrix; d k is the scaling factor of the attention mechanism.

[0064] Calculation of the attention weight matrix A' after time adjustment:

[0065] A′=A+β·TemporalAttention(H);

[0066] Among them, β is the dynamically adjusted weight coefficient, which is obtained by the sigmoid function.

[0067] TemporalAttention (H) is the output of the temporal attention mechanism, which accepts the user's historical behavior sequence H as input and calculates an attention matrix that reflects the temporal dynamics of the user's interest. The specific implementation of this function may involve the following steps: a. Encode each element in the historical behavior sequence H to obtain a series of vector representations. b. Use the self-attention mechanism or other temporal models (such as LSTM, GRU) to capture the temporal dependencies in the behavior sequence. c. Compare the processed temporal information with the current user interests and product features to generate an attention matrix with the same dimension as A.

[0068] Preferably, a dynamic weight adjustment mechanism is introduced to dynamically adjust the β value based on the correlation between the user's current status and historical behavior:

[0069]

[0070] Among them, W β is a learnable weight matrix, || represents a vector concatenation operation, and avg(H) is the average representation of the historical behavior sequence.

[0071] S2.4: If the value of an element in the attention weight matrix A' is lower than the sparsification threshold ε, the element is set to 0 to obtain the sparse attention weight matrix A'' to reduce the impact of noise and improve the efficiency of subsequent calculations.

[0072] Specifically, the sparse operation:

[0073]

[0074] The sparsification threshold ε is adjusted through an adaptive threshold mechanism:

[0075] ε=μ·meon(A′)×σ·std(A′);

[0076] Among them, μ and σ are adjustable hyperparameters, mean(A') and std(A') are the mean and standard deviation of A', respectively.

[0077] S2.5: Based on the sparse attention weight matrix A', the user-item matching score S is calculated; when the user-item matching score S exceeds the matching score threshold τ, the corresponding item is added to the initial candidate item set C; if the size of the initial candidate item set C reaches a preset number or all items have been evaluated, the user-item matching score S in the initial candidate item set C is sorted in descending order, and the sorted initial candidate item set C is output as the final candidate item set.

[0078] Specifically, the user-item matching score S is expressed by the following formula:

[0079]

[0080] Among them, S i is the user-item matching score of the ith item, which theoretically ranges from (0,2) but usually falls around (0,1) in practice, depending on the input value; i is the index of the item, indicating which item's user-item matching score is being calculated; d is the total dimension of the feature, which is equal to the sum of the item category dimension k, the brand dimension l, and the price range dimension m; j is the index of the feature dimension, from 1 to d; A″ ij is the weight value of the jth feature dimension of the ith product in the sparse attention weight matrix A”, reflecting the user’s attention to this specific feature; jis the value of the jth feature dimension of the ith product in the product feature matrix G, indicating the attribute of the product on this feature dimension; zi is the dot product sum of the attention weight and the product feature, which comprehensively considers the matching degree between the user interest (reflected by the attention weight) and the product feature;

[0081]

[0082] f(A″ ij , g ij )=ReLU(A″ ij +g ij -|A″ ij -g ij |);

[0083] Among them, f(A″ ij , g ij ) is a nonlinear function used to capture the interaction between attention weights and product features. When A” ij and g ij When they are close, the function outputs the larger value; when they are very different, it outputs the smaller value or 0.

[0084] The calculation of user-item matching score incorporates multiple advanced machine learning concepts and has significant advantages over traditional linear weighted models. First, by introducing the sigmoid function and the hyperbolic tangent function, the expression cleverly transforms the linear combination z i Mapping to a nonlinear space effectively captures the complex nonlinear relationship between user interests and product features. This nonlinear mapping can more accurately simulate actual user behavior patterns and significantly improve the accuracy and generalization ability of the recommendation system compared to simple linear models. ij , g ij ) introduces the concepts of attention mechanism and feature interaction, which is an important improvement over the existing technology. Specifically, the nonlinear function f(A″) ij , g ij ) not only considers the user's attention to a specific feature (through A" ij This design enables the model to adaptively adjust the importance of different features and capture the potential correlation between features, thereby greatly improving its expressiveness and recommendation accuracy while keeping the model simple. Compared with traditional methods, this dynamic feature interaction mechanism can better adapt to the personalized needs of different users and the unique attributes of different products, further improving the performance of the recommendation system and user satisfaction.

[0085] S3: Apply a dynamic context-aware reinforcement learning model to re-rank the candidate product set based on real-time user interaction behavior and environmental factors, and output the final recommendation results.

[0086] S3.1: Construct the state space, combining the user's multimodal interest feature vector, product feature matrix, historical behavior sequence H, current time and device type to form a state vector.

[0087] S3.2: Based on the user multimodal interest feature vector obtained in S1 and the user-item matching score calculated in S2, an action space is designed, where each action corresponds to a ranking method for items in the candidate item set C.

[0088] S3.3: A state encoder with a multi-head attention mechanism is used to map the state vector into a low-dimensional representation, where the number of attention heads matches the dimension of the user’s multimodal interest feature vector.

[0089] S3.4: Use the calculation formula of the user-item matching score in S2 as part of the reward function, and design a reinforcement learning reward function that comprehensively considers the item matching degree and user real-time feedback.

[0090] S3.5: Use a policy network to select the optimal action based on the low-dimensional representation and reward function.

[0091] S3.6: Reorder the candidate product set according to the optimal action, generate a sorted product list, and select the top K products as the final recommendation results to output to the user.

[0092] It should be noted that users’ real-time interactive behaviors are mainly reflected in:

[0093] S3.4: The reinforcement learning reward function takes into account real-time user feedback.

[0094] S3.5: The policy network makes decisions based on the current state, implicitly responding to the user's real-time behavior.

[0095] Environmental factors are mainly reflected in:

[0096] S3.1: The state vector contains the current time and device type.

[0097] S3.2: The design of the action space takes into account the set of candidate items.

[0098] In summary, the present invention realizes the comprehensive capture and accurate modeling of user interests by combining multimodal interest feature extraction, dynamic context perception and reinforcement learning technology. The present invention not only takes into account the user's historical behavior and long-term preferences, but also can respond to the user's immediate needs and environmental changes in real time, thereby performing intelligent reordering based on the initial set of candidate products. This innovative combination significantly improves the accuracy, personalization and adaptability of the recommendation system, and effectively solves the problem that traditional recommendation methods perform poorly in dynamic environments. At the same time, by introducing multi-head attention mechanisms and reinforcement learning strategies, the present invention can better capture the complex nonlinear relationship between user interests and product features, further improving the recommendation effect. Ultimately, the present invention can not only provide users with more accurate and personalized product recommendations, but also improve user satisfaction and the commercial value of the platform, which is of great significance to improving the user experience and operational efficiency of e-commerce platforms.

[0099] Embodiment 2 is an embodiment of the present invention, which provides an e-commerce platform product recommendation system based on user preference analysis, including: a construction module, used to construct a user multimodal interest feature vector; a generation module, used to use a cross-modal attention mechanism to match the user multimodal interest feature vector with the multimodal attribute information of the product to generate a set of candidate products; an output module, used to apply a dynamic context-aware reinforcement learning model, re-sort the candidate product set according to the user's real-time interactive behavior and environmental factors, and output the final recommendation result.

[0100] Example 3, reference Figure 2 , is an embodiment of the present invention, which is different from the previous embodiment in that: if the function is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk and other media that can store program codes.

[0101] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as an ordered list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by an instruction execution system, device or apparatus (such as a computer-based system, a system including a processor, or other system that can fetch instructions from an instruction execution system, device or apparatus and execute instructions), or in conjunction with such instruction execution systems, devices or apparatuses. For the purposes of this specification, "computer-readable medium" can be any device that can contain, store, communicate, propagate or transmit a program for use by an instruction execution system, device or apparatus, or in conjunction with such instruction execution systems, devices or apparatuses.

[0102] More specific examples of computer-readable media (a non-exhaustive list) include the following: an electrical connection with one or more wires (electronic device), a portable computer disk case (magnetic device), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), an optical fiber device, and a portable compact disk read-only memory (CDROM). In addition, the computer-readable medium may even be a paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, deciphering or, if necessary, processing in another suitable manner, and then stored in a computer memory.

[0103] It should be understood that the various parts of the present invention can be implemented by hardware, software, firmware or a combination thereof. In the above-mentioned embodiments, a plurality of steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, it can be implemented by any one of the following technologies known in the art or their combination: a discrete logic circuit having a logic gate circuit for implementing a logic function for a data signal, a dedicated integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0104] Example 4 is an embodiment of the present invention, which provides an e-commerce platform product recommendation method based on user preference analysis. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through economic benefit calculation and simulation experiments.

[0105] This experiment is conducted on a real dataset of a large e-commerce platform, which contains interaction data of 10 million users and 5 million products over a time span of 6 months. This experiment divides the dataset into a training set (first 5 months) and a test set (last 1 month) in chronological order. For each method, this experiment uses the training set for model training and parameter optimization, and then evaluates the model performance on the test set.

[0106] To ensure a fair comparison, this experiment uses the same preprocessing steps and evaluation metrics for all methods. During the experiment, this experiment simulates a real online recommendation scenario, including factors such as user real-time behavior and environmental changes. For the present invention, this experiment pays special attention to its ability to handle multimodal data, dynamic context, and real-time interaction. The experiment was repeated 5 times and the average value was taken as the final result to reduce the impact of randomness.

[0107] Table 1 Comparison of recommended performance indicators

[0108] method Accuracy (%) Recall rate (%) F1 score NDCG MAP Collaborative filtering 72.3 68.5 0.703 0.681 0.625 Content basic recommendation 75.1 70.2 0.726 0.695 0.648 Deep Collaborative Filtering 78.6 73.8 0.761 0.729 0.683 Attention Mechanism Model 81.2 76.5 0.788 0.752 0.715 The present invention 85.7 82.3 0.839 0.801 0.769

[0109] As shown in Table 1, the present invention significantly outperforms other methods in all indicators, with precision and recall reaching 85.7% and 82.3% respectively, which are 4.5 and 5.8 percentage points higher than the second best attention mechanism model. The advantages of F1 score (0.839), NDCG (0.801) and MAP (0.769) are even more obvious, increasing by 6.5%, 6.5% and 7.6% respectively. These results show that the present invention can not only more accurately identify the products that users are interested in, but also provide higher quality ranking results. In particular, the significant improvement in NDCG and MAP reflects the excellent performance of the present invention in generating personalized ranking lists, which is crucial to improving user experience.

[0110] Table 2: Comparison of user interaction and business indicators

[0111]

[0112] As shown in Table 2, the advantages of the present invention in these indicators are more prominent, with the click-through rate (6.8%) and conversion rate (3.5%) being 21.4% and 20.7% higher than the closest competitors, respectively. More noteworthy is that the present invention significantly prolonged the average user session duration (15.6 minutes), improved the user retention rate (61.2%), and brought higher average revenue per user (112.3 yuan). The significant improvement in these indicators proves that the present invention can not only provide more accurate recommendations, but also effectively improve user satisfaction and the commercial value of the platform. In particular, the substantial increase in user retention rate and average revenue reflects the advantages of the present invention in long-term user value mining.

[0113] Table 3: Comparison of recommendation system’s ability to handle challenging problems

[0114]

[0115] As shown in Table 3, the present invention performs particularly well in cold start performance (0.749) and long-tail product coverage (28.6%), which are 20.6% and 25.4% higher than the closest competitors, respectively. This shows that the present invention can better handle the recommendation problems of new users and new products, and can effectively tap the potential of long-tail products. In terms of real-time response time, although the present invention (89ms) is not the fastest, considering its complexity, this response speed is very satisfactory. In addition, the present invention also performs best in recommendation diversity (0.812) and noise resistance (0.785), which reflects its advantages in providing diversified recommendations and processing noisy data. These results fully demonstrate the comprehensiveness and effectiveness of the present invention in dealing with the complex challenges of recommendation systems.

[0116] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A method for recommending products on an e-commerce platform based on user preference analysis, characterized in that: include: Constructing a user multimodal interest feature vector; the multimodal interest feature vector integrates the user's text search history, image browsing preference, audio consumption habits, and video interaction behavior; Using a cross-modal attention mechanism, the user's multimodal interest feature vector is matched with the multimodal attribute information of the product to generate a candidate product set; Applying a dynamic context-aware reinforcement learning model, re-ranking the candidate product set according to the user's real-time interactive behavior and environmental factors, and outputting a final recommendation result; Constructing a user's multimodal interest feature vector includes the following steps: Extract user browsing history, purchase history, and search history from the e-commerce platform database; Clean the user's historical browsing records, purchase records, and search records, remove duplicate data and invalid data, and classify the cleaned user's historical browsing records, purchase records, and search records according to product category, brand, and price range; Based on the classified historical browsing records, purchase records, and search records, the user's preference weights for each product category, brand, and price range are calculated to generate a user's multimodal interest feature vector; The cross-modal attention mechanism is used to match the user's multimodal interest feature vector with the multimodal attribute information of the product to generate a candidate product set, including the following steps: Receiving the user's multimodal interest feature vector; Based on the multimodal attribute information of commodities in the e-commerce platform database, a commodity feature matrix is ​​constructed; at the same time, the historical behavior sequence of user interactions in the historical time is obtained; When the user's multimodal interest feature vector, product feature matrix and historical behavior sequence are input, an initial attention weight matrix is ​​calculated through a dynamic weight cross-modal attention network; the weights in the initial attention weight matrix are adjusted based on the historical behavior sequence using a temporal attention mechanism to obtain an attention weight matrix; If the value of an element in the attention weight matrix is ​​lower than the sparsification threshold, the element is set to 0 to obtain a sparse attention weight matrix; Based on the sparse attention weight matrix, the user-item matching score is calculated: when the user-item matching score exceeds the matching score threshold τ, the corresponding item is added to the initial candidate item set; if the size of the initial candidate item set reaches a preset number or all items have been evaluated, the items in the initial candidate item set are sorted in descending order according to the user-item matching score, and the sorted initial candidate item set is output as the final candidate item set; The user-item matching score is expressed by the following formula: ； in, It is i User-item matching score for each item; i It is the index of the commodity; d is the total dimension of the feature, equal to the product category dimension k , Brand dimension and price range dimension m The sum of j is the index of the feature dimension, from 1 to d ; is the first in the sparse attention weight matrix A'' i Product j The weight value of feature dimensions; is the first i Product j The value of the feature dimension; It is the dot product sum of the attention weight and the product features; ； ； in, is a nonlinear function used to capture the interaction between attention weights and product features.

2. The method for recommending products on an e-commerce platform based on user preference analysis as claimed in claim 1, characterized in that: The dynamic weight cross-modal attention network dynamically captures the relationship between user interests and product features through the attention mechanism, while considering the interactions between product categories, brands, and price ranges.

3. The method for recommending products on an e-commerce platform based on user preference analysis as claimed in claim 2, characterized in that: The reinforcement learning model for applying dynamic context awareness includes the following steps: constructing a state space, combining a user's multimodal interest feature vector, a commodity feature matrix, a historical behavior sequence H, a current time, and a device type to form a state vector; Designing an action space based on the user multimodal interest feature vector and the user-product matching score; A state encoder with a multi-head attention mechanism is used to map the state vector into a low-dimensional representation, where the number of attention heads matches the dimension of the user’s multimodal interest feature vector. Using the user-item matching score, designing a reinforcement learning reward function that comprehensively considers item matching and user real-time feedback; Based on the low-dimensional representation and the reinforcement learning reward function, a policy network is used to select the optimal action.

4. The method for recommending products on an e-commerce platform based on user preference analysis as claimed in claim 3, characterized in that: The candidate product set is reordered according to the optimal action to generate a sorted product list, and the top K products are selected as the final recommendation results and output to the user.

5. A system using the e-commerce platform product recommendation method based on user preference analysis as described in any one of claims 1 to 4, characterized in that: include: Construction module, used to construct user multimodal interest feature vector; A generation module, used to match the user's multimodal interest feature vector with the multimodal attribute information of the product by using a cross-modal attention mechanism to generate a candidate product set; The output module is used to apply a dynamic context-aware reinforcement learning model to reorder the candidate product set according to the user's real-time interactive behavior and environmental factors, and output a final recommendation result.

6. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, the steps of the e-commerce platform product recommendation method based on user preference analysis described in any one of claims 1 to 4 are implemented.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method for recommending products on an e-commerce platform based on user preference analysis described in any one of claims 1 to 4 are implemented.

Images