Commodity recommendation method and device, equipment, storage medium and product

Abstract

The invention discloses a commodity recommendation method and device, equipment, a storage medium and a product, and relates to the technical field of commodity recommendation, and the method comprises the steps: constructing a user portrait feature vector according to the historical behavior data of a target user; screening out a candidate commodity set from a target commodity library based on the user portrait feature vector; determining an attention index of the target user, and determining a target score of the attention index corresponding to each candidate commodity in the candidate commodity set; inputting the target score into a regression model to obtain a comprehensive score of the candidate commodity; and sorting all candidate commodities in the candidate commodity set according to the comprehensive score to obtain a sorting result, and recommending related commodities to the target user according to the sorting result. According to the invention, more appropriate commodities can be recommended to the user.

Description

Technical Field

[0001] This application relates to the field of product recommendation technology, and in particular to product recommendation methods, apparatus, equipment, storage media, and products. Background Technology

[0002] With the rapid development of e-commerce, personalized product recommendations have become a key technology for improving user experience. Traditional recommendation schemes achieve product matching through single-dimensional recommendation strategies, such as considering only purchase history or product tags. This fails to comprehensively capture the multi-layered needs of users, resulting in insufficient accuracy of recommendation results.

[0003] The above content is only used to help understand the technical solution of this application and does not represent an admission that the above content is prior art. Summary of the Invention

[0004] The main objective of this application is to provide a product recommendation method, apparatus, device, storage medium, and product, aiming to solve the technical problem that traditional recommendation methods cannot fully capture the multi-level needs of users, resulting in insufficient accuracy of recommendation results.

[0005] To achieve the above objectives, this application proposes a product recommendation method, which includes:

[0006] Construct user profile feature vectors based on the target users' historical behavior data;

[0007] Based on the user profile feature vector, a candidate product set is selected from the target product library;

[0008] Determine the attention metrics of the target user, and determine the target score for the attention metrics corresponding to each candidate product in the candidate product set;

[0009] The target score is input into the regression model to obtain the comprehensive score of the candidate products;

[0010] All candidate products in the candidate product set are sorted according to the comprehensive score to obtain the sorting result, and relevant products are recommended to the target user based on the sorting result.

[0011] In one embodiment, the step of constructing a user profile feature vector based on the target user's historical behavior data includes:

[0012] Obtain the target user's historical behavior data;

[0013] Consumption characteristics, behavioral characteristics, and social characteristics are extracted from the historical behavioral data;

[0014] First and second cross features are extracted from the consumption features, behavioral features, and social features;

[0015] The first cross feature and the second cross feature are subjected to higher-order feature cross processing to generate higher-order combined features;

[0016] Principal component analysis is performed on the higher-order combined features to reduce their dimensionality, resulting in low-dimensional dense features;

[0017] Based on the low-dimensional dense features, the consumption features, the behavioral features, and the social features, a user profile feature vector is constructed.

[0018] In one embodiment, the step of selecting a candidate product set from the target product library based on the user profile feature vector includes:

[0019] Determine the product feature vector of each product in the target product library;

[0020] The user profile feature vector and the product feature vector are input into the deep matching model. The user profile feature vector is encoded by the user tower of the deep matching model to obtain the user representation vector. The product feature vector is encoded by the product tower of the deep matching model to obtain the product representation vector.

[0021] Determine the multigranular similarity between the user representation vector and the product representation vector;

[0022] Based on the multi-granular similarity, a candidate product set is selected from the target product library.

[0023] In one embodiment, the step of determining the multi-granularity similarity between the user representation vector and the product representation vector includes:

[0024] The cosine similarity between the user representation vector and the product representation vector is used as the coarse-grained similarity.

[0025] Based on the attention mechanism, the user representation vector and the product representation vector are divided into multiple sub-vector groups, and the granular similarity is determined according to the similarity between the sub-vector groups. Each sub-vector group includes user representation word vectors and product representation word vectors.

[0026] Fine-grained similarity is obtained by performing element-wise interaction operations on the user representation vector and the product representation vector.

[0027] Based on the coarse-grained similarity, the medium-grained similarity, and the fine-grained similarity, multi-grained similarity is determined.

[0028] In one embodiment, the step of performing element-wise interaction operations on the user representation vector with the product representation vector to obtain fine-grained similarity includes:

[0029] Based on the user representation vector and the product representation vector, an element-by-element interaction vector is generated.

[0030] The element-wise interaction vector, the user representation vector, and the product representation vector are concatenated to obtain an enhanced feature vector.

[0031] Fine-grained similarity is obtained by performing a nonlinear transformation on the enhanced feature vector using a multilayer perceptron.

[0032] In one embodiment, the step of determining the target user's attention metrics includes:

[0033] Obtain explicit and implicit feedback data from the target user;

[0034] Extract the target user's specific attention indicators from the displayed feedback data;

[0035] Extract the potential attention indicators of the target user from the implicit feedback data;

[0036] Based on the explicitly concerned indicators and the potentially concerned indicators, an initial indicator set is generated;

[0037] The timeliness weight of each indicator in the initial indicator set is determined based on the user's recent behavior data, and the importance weight of each indicator in the initial indicator set is learned through an attention mechanism.

[0038] Based on the timeliness weight and the importance weight, the dynamic weight evaluation results of each indicator in the initial indicator set are determined;

[0039] Based on the dynamic weight evaluation results, an optimized index set is determined according to the initial index set.

[0040] The target user's focus metrics are determined from the set of optimization metrics.

[0041] Furthermore, to achieve the above objectives, this application also proposes a product recommendation device, which includes:

[0042] The module is used to construct user profile feature vectors based on the target user's historical behavior data;

[0043] The filtering module is used to filter out a set of candidate products from the target product library based on the user profile feature vector;

[0044] The determination module is used to determine the attention indicators of the target user and to determine the target score of the attention indicators corresponding to each candidate product in the candidate product set.

[0045] The input module is used to input the target score into the regression model to obtain the comprehensive score of the candidate products;

[0046] The sorting module is used to sort all candidate products in the candidate product set according to the comprehensive score, obtain the sorting result, and recommend relevant products to the target user based on the sorting result.

[0047] In addition, to achieve the above objectives, this application also proposes a product recommendation device, the device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the product recommendation method as described above.

[0048] In addition, to achieve the above objectives, this application also proposes a storage medium, which is a computer-readable storage medium, on which a computer program is stored, and which, when executed by a processor, implements the steps of the product recommendation method described above.

[0049] In addition, to achieve the above objectives, this application also provides a computer program product, which includes a computer program that, when executed by a processor, implements the steps of the product recommendation method described above.

[0050] One or more technical solutions proposed in this application have at least the following technical effects:

[0051] The product recommendation method, apparatus, device, storage medium, and product proposed in this application construct a user profile feature vector based on the historical behavior data of the target user; based on the user profile feature vector, a candidate product set is selected from the target product library; the target user's attention indicators are determined, and the target score of the attention indicator corresponding to each candidate product in the candidate product set is determined; the target score is input into a regression model to obtain a comprehensive score of the candidate products; all candidate products in the candidate product set are sorted according to the comprehensive score to obtain a sorting result, and relevant products are recommended to the target user based on the sorting result. This invention addresses the technical problem of insufficient accuracy in traditional recommendation methods due to their inability to fully capture the multi-layered needs of users. Compared to existing technologies, this application first constructs a user profile feature vector using historical user behavior data and performs deep matching with the product feature vectors in the target product library to filter out a candidate product set. Based on this, the application further scores the attention indicators corresponding to each candidate product in the candidate product set. Then, it uses a regression model to comprehensively consider the scores of multiple attention indicators to obtain a more accurate comprehensive score for the candidate products. Finally, the candidate products are ranked according to the comprehensive score and recommended to users, thereby ensuring that the recommendation results better meet the personalized needs of users and significantly improving the accuracy of recommendations and user experience. Attached Figure Description

[0052] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0053] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0054] Figure 1 This is a flowchart illustrating an embodiment of the product recommendation method of this application.

[0055] Figure 2 This is a schematic diagram of single-model multi-task scoring fusion provided in Embodiment 1 of the product recommendation method of this application;

[0056] Figure 3 This is a schematic diagram of multi-model single-task scoring fusion provided in Embodiment 1 of the product recommendation method of this application;

[0057] Figure 4 This is a flowchart of the regression model update provided in Embodiment 1 of the product recommendation method of this application;

[0058] Figure 5This is a flowchart illustrating Embodiment 2 of the product recommendation method of this application;

[0059] Figure 6 This is a schematic diagram of the module structure of the product recommendation device according to an embodiment of this application;

[0060] Figure 7 This is a schematic diagram of the device structure of the hardware operating environment involved in the product recommendation method in this application embodiment.

[0061] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0062] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0063] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0064] The main solution of this application embodiment is as follows: construct a user profile feature vector based on the target user's historical behavior data; based on the user profile feature vector, select a set of candidate products from the target product library; determine the target user's attention indicators and determine the target score of the attention indicators corresponding to each candidate product in the candidate product set; input the target score into a regression model to obtain the comprehensive score of the candidate products; sort all candidate products in the candidate product set according to the comprehensive score to obtain the sorting result, and recommend relevant products to the target user according to the sorting result.

[0065] As can be seen from the above embodiments, the product recommendation method, apparatus, device, storage medium, and product proposed in this application construct a user profile feature vector based on the target user's historical behavior data; based on the user profile feature vector, select a candidate product set from the target product library; determine the target user's attention indicators and determine the target score of the attention indicators corresponding to each candidate product in the candidate product set; input the target score into a regression model to obtain a comprehensive score for the candidate products; sort all candidate products in the candidate product set according to the comprehensive score to obtain a sorting result, and recommend relevant products to the target user based on the sorting result. This invention addresses the technical problem of insufficient accuracy in traditional recommendation methods due to their inability to fully capture the multi-layered needs of users. Compared to existing technologies, this application first constructs a user profile feature vector using historical user behavior data and performs deep matching with the product feature vectors in the target product library to filter out a candidate product set. Based on this, the application further scores the attention indicators corresponding to each candidate product in the candidate product set. Then, it uses a regression model to comprehensively consider the scores of multiple attention indicators to obtain a more accurate comprehensive score for the candidate products. Finally, the candidate products are ranked according to the comprehensive score and recommended to users, thereby ensuring that the recommendation results better meet the personalized needs of users and significantly improving the accuracy of recommendations and user experience.

[0066] It should be noted that the executing entity in this embodiment can be a computing service device with data processing, network communication, and program execution functions, such as a tablet computer, personal computer, or mobile phone, or an electronic device or product recommendation device capable of performing the above functions. The following description uses a product recommendation device as an example to illustrate this embodiment and the subsequent embodiments.

[0067] Based on this, the embodiments of this application provide a product recommendation method, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the product recommendation method of this application.

[0068] In this embodiment, the product recommendation method includes steps S10 to S50:

[0069] Step S10: Construct a user profile feature vector based on the target user's historical behavior data;

[0070] It should be noted that historical behavioral data includes users' browsing history (such as visits to product detail pages, search keywords), purchase history (such as order data, repurchase cycle), favorites / add-to-cart behavior (such as adding products to favorites, adding items to the shopping cart), and interaction behavior (such as comments, likes, and shares). User profile feature vectors are low-dimensional dense vectors obtained through feature extraction and dimensionality reduction of user behavioral data, used to represent user preferences. By constructing user profile feature vectors, complex user behavioral data can be transformed into quantifiable feature representations for subsequent similarity calculations and recommendation matching.

[0071] In one feasible implementation, the step of constructing a user profile feature vector based on the target user's historical behavior data includes: acquiring the target user's historical behavior data; extracting consumption features, behavioral features, and social features from the historical behavior data; mining a first cross feature and a second cross feature from the consumption features, behavioral features, and social features; performing high-order feature cross processing on the first cross feature and the second cross feature to generate high-order combined features; performing principal component analysis to reduce the dimensionality of the high-order combined features to obtain low-dimensional dense features; and constructing a user profile feature vector based on the low-dimensional dense features, the consumption features, the behavioral features, and the social features.

[0072] It should be noted that the consumption characteristics include: purchase frequency, average order value, and category preference (e.g., 3C products account for 60%); behavioral characteristics include: click-through rate, page dwell time, and TF-IDF of search keywords; and social characteristics include: frequency of interaction with friends and shared product categories.

[0073] It should be noted that factorization machine (FM) can be used for second-order feature cross-processing. Specifically, latent vector inner product calculation is performed on consumption features and behavioral features to uncover the "high consumption + nighttime activity" cross-feature (i.e., the first cross-feature); latent vector inner product calculation is performed on behavioral features and social features to uncover the "high-frequency search + strong social influence" cross-feature (i.e., the second cross-feature). Furthermore, a three-layer deep neural network combined with a multi-head attention mechanism is used to perform high-order feature cross-processing on the cross-features, generating a 64-dimensional high-order combined feature. Then, principal component analysis (PCA) with 95% variance retention is used to reduce the dimensionality to 32-dimensional low-dimensional dense features.

[0074] Step S20: Based on the user profile feature vector, select a set of candidate products from the target product library;

[0075] It's important to note that the target product library refers to a collection containing all products to be recommended on the platform. Each product has structured features (such as category, price, and sales volume) and unstructured features (such as images and descriptive text). A preliminary set of Top-N products (typically N = 100-500) can be selected as the candidate product set by matching user profile feature vectors with product feature vectors from the target product library. This allows for the rapid selection of a small number of potentially relevant products from a massive product library, avoiding full-scale computation.

[0076] Step S30: Determine the attention indicators of the target user and determine the target score of the attention indicators corresponding to each candidate product in the candidate product set;

[0077] It's important to clarify that the key performance indicators (KPIs) refer to the multi-dimensional evaluation factors that users prioritize when making product decisions. These include explicit KPIs (such as price, ratings, and sales volume, which can be directly quantified) and implicit KPIs (such as social influence and new product preference, which require behavioral analysis to uncover). In practice, the first step is to establish a KPI scoring model by analyzing the historical behavioral data of all users on the platform (including but not limited to clickstream, purchase records, and reviews). Explicit KPIs are directly quantified using statistical analysis methods (such as purchase price distribution and rating click-through rates), while implicit KPIs are extracted using deep learning models (such as LSTM for behavioral sequence analysis and GNN for social relationship mining). Secondly, for each candidate product, the target score for each KPI is calculated: the price KPI uses a dynamic price sensitivity model, combining historical user spending patterns and real-time market data to calculate the matching degree; the quality KPI is evaluated using a multi-dimensional weighted assessment of brand level, material certification, and other factors; and the social KPI is based on graph neural network analysis of the user's influence on social networks. All scores are normalized (0-1 range) and a time decay mechanism is established (70% weighting of behavior in the most recent 30 days), ultimately forming a complete product-indicator score matrix, providing accurate multi-dimensional data support for subsequent comprehensive recommendations.

[0078] It should be noted that, as Figure 2 as well as Figure 3 As shown, when there are multiple objectives related to the focus indicators (i.e., multiple focus indicators), multiple models or a single multi-task model (i.e., multiple task models or a single MMoE model) can be used to score the focus indicators corresponding to the candidate products, thereby obtaining the target scores for the focus indicators corresponding to the candidate products. By establishing a dynamic multi-dimensional indicator scoring system, it is possible to more accurately reflect the real-time changes in user preferences and solve the recommendation bias problem caused by the single indicator in traditional recommendation methods.

[0079] In one feasible implementation, the step of determining the target user's attention metrics includes: acquiring explicit feedback data and implicit feedback data of the target user; extracting explicit attention metrics of the target user from the explicit feedback data; extracting potential attention metrics of the target user from the implicit feedback data; generating an initial metric set based on the explicit attention metrics and the potential attention metrics; determining the timeliness weight of each metric in the initial metric set based on the user's recent behavior data, and learning the importance weight of each metric in the initial metric set through an attention mechanism; determining the dynamic weight evaluation result of each metric in the initial metric set based on the timeliness weight and the importance weight; determining an optimized metric set based on the dynamic weight evaluation result and the initial metric set; and determining the target user's attention metrics from the optimized metric set.

[0080] It should be noted that explicit feedback data includes, but is not limited to: user rating data, product collection records, questionnaire results, and customer service consultation records; implicit feedback data includes, but is not limited to: page dwell time, clickstream data, search keywords, add-to-cart behavior, and sharing behavior.

[0081] Understandably, during the indicator extraction phase, explicitly relevant indicators (such as price sensitivity) are directly extracted from explicit data using a rule engine (Drools), while potential indicators (such as social influence) are mined using deep learning methods (e.g., using Bi-LSTM to analyze user behavior sequence patterns). After the initial indicator set is generated, the system establishes a two-layer dynamic evaluation system: the timeliness weight calculation module processes user behavior data from the past 30 days based on an exponential decay function (weight = base^(-λt), where base is 0.9 and λ = 0.05) to ensure that recent behavior receives higher weight; the attention mechanism uses a 3-layer Transformer structure (128 hidden layer dimensions, 8 attention heads) to learn the complex relationships between indicators, and visualizes the distribution of indicator importance through a multi-head attention weight matrix. In the dynamic weight fusion phase, a gating mechanism (Gated LinearUnit) is introduced to adaptively fuse timeliness weights and attention weights. The final optimized indicator set is then double-validated to generate the final focus indicators. In particular, a real-time monitoring mechanism has been established to track the validity period, confidence level and fluctuation of each indicator in real time, and to automatically trigger an early warning mechanism when an indicator is abnormal (such as a sudden 30% drop in weight).

[0082] It should be noted that the steps for determining the target user's focus metrics are further subdivided into a dual-track management system of static and dynamic metrics. Static metrics (such as brand loyalty, basic category preferences, and other long-term stable characteristics) are calculated by analyzing user historical data for more than 6 months using a moving average algorithm (window size 180 days), and are incrementally updated monthly, with a stability verification mechanism (fluctuation threshold ±15%). Dynamic metrics (such as promotion sensitivity, seasonal preferences, and other short-term changing characteristics) are based on behavioral data from the past 30 days, processed by a real-time calculation engine (such as Flink), and dynamically updated using a dual time decay factor model (70% weight for the recent 7 days and 30% for 8-30 days), with an update frequency of once per hour. The system establishes a metric status transition mechanism: when a dynamic metric remains stable for 30 consecutive days and has a confidence level ≥90%, it can be upgraded to a static metric after manual review; conversely, when a static metric fluctuates continuously beyond the threshold, it is automatically downgraded to a dynamic metric for reassessment. In terms of technical implementation, static metrics are stored in a relational database (such as MySQL) to ensure strong consistency, while dynamic metrics utilize a distributed cache (such as Redis) to ensure high-concurrency access performance. The system also features a metric health assessment matrix, monitoring three dimensions: timeliness (data freshness), stability (variance coefficient), and significance (SHAP value). A four-level early warning mechanism is implemented: a yellow alert (logged) is triggered when the metric health falls below 60%, an orange alert (automatic demotion) is triggered when it falls below 40%, and a red alert (suspension of use and notification to operations personnel) is triggered when it falls below 20%. In practical testing on an e-commerce platform, this system increased the exposure rate of long-tail products by 25%, and through the anchoring effect of static metrics, improved the recommendation accuracy during the cold start phase for new users by 18%.

[0083] Step S40: Input the target score into the regression model to obtain the comprehensive score of the candidate product;

[0084] It should be noted that regression models (such as linear regression and XGBoost) are used to integrate multiple indicator scores and output a comprehensive score. Regression models can learn the non-linear relationships between indicators, avoiding the subjectivity of manual weighting. Since product metrics change with business indicators, the scoring criteria will also change accordingly. Figure 4 As shown, the regression model can be trained and updated using the model scoring criteria log, thus enabling the regression model to adapt to various scenarios and achieve dynamic updates.

[0085] In practical implementation, we focus on two key metrics: clicks and conversions. The formula for calculating the comprehensive score of candidate products by inputting these target scores into the regression model is as follows:

[0086] Y = β0 + β1X1 + β2X2 + ε

[0087] In the formula, Y represents the overall score, X1 represents the target score corresponding to the click index, X2 represents the target score corresponding to the conversion index, β0 is the intercept term, ε is the error term (representing the random error that the regression model cannot explain), β1 represents the fixed parameter corresponding to the click index, and β2 represents the fixed parameter corresponding to the conversion index.

[0088] Step S50: Sort all candidate products in the candidate product set according to the comprehensive score to obtain the sorting result, and recommend relevant products to the target user based on the sorting result.

[0089] It should be noted that a smart sorting algorithm can be used to transform candidate products into a final recommendation list, and then relevant products can be recommended to target users based on the final recommendation list. Specifically: First, all candidate products in the candidate product set are sorted based on comprehensive scores using a fast sorting algorithm (time complexity O(nlog n)) to arrange all candidate products in descending order of scores; second, a diversity control mechanism is introduced, using the MMR (Maximal Marginal Relevance) algorithm to balance the accuracy and diversity of the recommendation results, where the similarity threshold is set to 0.7 to ensure that similar products do not exceed 30% of the total number of recommendations; then, business rule filtering is performed to exclude products that do not meet the conditions, such as insufficient inventory or being discontinued; finally, fine-tuning is performed based on real-time contextual features (such as the user's current geographical location and device type) to generate a Top-K (usually K=10-20) final recommendation list.

[0090] It should be noted that, at the display level, an A / B testing framework is used to dynamically optimize the recommended placement and display format. For example, user reviews are prioritized for high-priced items, and price information is highlighted for promotional items. This ranking mechanism, by balancing accuracy, diversity, and business needs, improves the NDCG (Normalized Discounted Cumulative Gain) metric by 35% compared to traditional single ranking methods, while also increasing user click-through rate by 22%.

[0091] This embodiment constructs a user profile feature vector based on the target user's historical behavior data; based on the user profile feature vector, a candidate product set is selected from the target product library; the target user's attention indicators are determined, and the target score for each candidate product in the candidate product set corresponding to the attention indicators is determined; the target score is input into a regression model to obtain a comprehensive score for the candidate products; all candidate products in the candidate product set are sorted according to the comprehensive score to obtain a sorting result, and relevant products are recommended to the target user based on the sorting result. This solves the technical problem of insufficient accuracy in traditional recommendation methods due to their inability to fully capture the multi-level needs of users. Compared with existing technologies, this application first constructs a user profile feature vector based on the user's historical behavior data and performs deep matching with the product feature vectors in the target product library to select a candidate product set. Based on this, this application further scores the attention indicators corresponding to each candidate product in the candidate product set, and then uses a regression model to comprehensively consider the scores of multiple attention indicators to obtain a more accurate comprehensive score for the candidate products. Finally, the candidate products are sorted according to the comprehensive score and recommended to the user, thereby ensuring that the recommendation results can better meet the user's personalized needs and significantly improving the accuracy of the recommendation and the user experience.

[0092] Based on the first embodiment of this application, in the second embodiment of this application, the content that is the same as or similar to the first embodiment described above can be referred to the above description, and will not be repeated hereafter. Based on this, please refer to... Figure 5 Step S20 also includes steps S201 to S204:

[0093] Step S201: Determine the product feature vector of each product in the target product library;

[0094] It should be noted that the product feature vector can include multi-dimensional information such as product category, price, brand, historical sales volume, and user reviews.

[0095] Step S202: Input the user profile feature vector and the product feature vector into the deep matching model, encode the user profile feature vector through the user tower of the deep matching model to obtain the user representation vector, and encode the product feature vector through the product tower of the deep matching model to obtain the product representation vector.

[0096] It should be noted that the deep matching model refers to a deep matching model that uses a dual-tower model (user tower and product tower). The user tower is used specifically to process user features, encoding user profile feature vectors into higher-level user representation vectors; the product tower is used specifically to process product features, encoding raw product feature vectors into higher-level product representation vectors.

[0097] Step S203: Determine the multi-granular similarity between the user representation vector and the product representation vector;

[0098] It should be noted that multi-granularity similarity is not a single similarity calculation, but rather a calculation of the degree of matching between users and products at multiple granularities / levels, such as global similarity, category-level similarity, feature-level similarity, and other different dimensions of matching. This multi-granularity method can more comprehensively evaluate the relationship between users and products.

[0099] In one feasible implementation, the step of determining the multi-granular similarity between the user representation vector and the product representation vector includes: using the cosine similarity between the user representation vector and the product representation vector as coarse-grained similarity; dividing the user representation vector and the product representation vector into multiple sub-vector groups based on an attention mechanism, and determining the medium-grained similarity based on the similarity between each group of sub-vectors, wherein each group of sub-vectors includes a user representation word vector and a product representation word vector; performing element-wise interactive operations on the user representation vector with respect to the product representation vector to obtain fine-grained similarity; and determining the multi-granular similarity based on the coarse-grained similarity, the medium-grained similarity, and the fine-grained similarity.

[0100] It's important to note that coarse-grained similarity refers to the cosine similarity between user representation vectors and product representation vectors, reflecting the directional consistency between users and products in the overall feature space. Medium-grained similarity can be calculated using word vector matching based on an attention mechanism (medium-grained similarity reveals the local matching of user interests and product features across different dimensions / aspects; the attention mechanism dynamically adjusts the importance of each dimension; it's more granular than global similarity and can discover matching of specific features). Specifically, the user representation vector is divided into multiple user sub-vectors using an attention mechanism, and the product representation vector is similarly divided into corresponding product sub-vectors. Then, the similarity between each group of user sub-vectors and product sub-vectors is calculated, and finally, the similarity results of all sub-vector groups are aggregated (i.e., medium-grained similarity). Fine-grained similarity is mainly calculated through element-wise interaction operations, which can capture the interaction relationships of each individual dimension in the feature vectors and discover very specific feature matching patterns.

[0101] Understandably, high coarse-grained similarity indicates that the product generally matches the user's interests; high medium-grained similarity indicates that the product matches the user's preferences in certain specific dimensions (such as brand and price); and high fine-grained similarity indicates that the product is more in line with the user's needs in terms of detailed features (such as color and size).

[0102] In practical implementation, a minimum threshold for coarse-grained similarity can be set. Only when a product's coarse-grained similarity exceeds this threshold is it included in the candidate product set. Coarse-grained similarity reflects the directional consistency between the user and the product in the overall feature space; therefore, a higher coarse-grained similarity means that the product generally matches the user's interests. A minimum threshold for medium-grained similarity can be set to filter products that match the user's preferences on certain specific dimensions (such as brand and price). Medium-grained similarity dynamically adjusts the importance of each dimension through an attention mechanism, revealing the local matching between user interests and product features on different dimensions. A minimum threshold for fine-grained similarity can be set to filter products that better meet the user's needs in detailed features (such as color and size). Fine-grained similarity captures the interaction relationships of each individual dimension in the feature vector through element-wise interaction operations, discovering very specific feature matching patterns.

[0103] In this embodiment, the candidate product set is selected by using multi-granular similarity. This method combines local features and fine-grained interactions to perform the selection while relying on global similarity, thereby avoiding the selection bias caused by a single similarity and affecting the selection results.

[0104] In one feasible implementation, the step of performing element-wise interaction operations on the user representation vector and the product representation vector to obtain fine-grained similarity includes: generating an element-wise interaction vector based on the user representation vector and the product representation vector; concatenating the element-wise interaction vector, the user representation vector, and the product representation vector to obtain an enhanced feature vector; and performing a nonlinear transformation on the enhanced feature vector using a multilayer perceptron to obtain fine-grained similarity.

[0105] In the specific implementation, the user representation vector u = (u1, u2, ..., u... n ) and the product representation vector v = (v1, v2, ..., v n Element-wise operations are performed to generate element-wise interactive vectors. Specifically, Hadamard product (i.e., element-wise multiplication), element-wise difference, or element-weighted summation can be used for interactive computation.

[0106] Step S204: Based on the multi-granular similarity, select a set of candidate products from the target product library.

[0107] In practice, some products can be initially screened based on the minimum thresholds for coarse-grained similarity, medium-grained similarity, and fine-grained similarity. Then, after determining the multi-grained similarity based on coarse-grained, medium-grained, and fine-grained similarity, candidate products can be screened from these initial products again.

[0108] This embodiment determines the product feature vectors of each product in the target product library; inputs the user profile feature vectors and product feature vectors into a deep matching model; encodes the user profile feature vectors using the user tower of the deep matching model to obtain user representation vectors; and encodes the product feature vectors using the product tower of the deep matching model to obtain product representation vectors; determines the multi-granular similarity between the user representation vectors and the product representation vectors; and selects a candidate product set from the target product library based on the multi-granular similarity. This application inputs the user profile feature vectors and product feature vectors into a deep matching model, using the model's user tower and product tower to encode the user profile and product features respectively, generating user representation vectors and product representation vectors. This dual-tower deep matching model can effectively mine the intrinsic features of users and products, encoding complex feature information into more representative representation vectors, thereby improving the accuracy and efficiency of feature matching. Furthermore, by determining the multi-granularity similarity between the user representation vector and the product representation vector, this embodiment breaks through the limitations of traditional single similarity calculation. Multi-granularity similarity calculates the matching degree between users and products at multiple granularities / levels, which can more comprehensively evaluate the relationship between users and products and avoid the bias that may be caused by single similarity calculation.

[0109] It should be noted that the above examples are only for understanding this application and do not constitute a limitation on the product recommendation method of this application. Any simple modifications based on this technical concept are within the protection scope of this application.

[0110] This application also provides a product recommendation device, please refer to... Figure 6 The product recommendation device includes:

[0111] Module 10 is used to construct user profile feature vectors based on the target user's historical behavior data;

[0112] The filtering module 20 is used to filter out a set of candidate products from the target product library based on the user profile feature vector;

[0113] The determining module 30 is used to determine the attention indicators of the target user and to determine the target score of the attention indicators corresponding to each candidate product in the candidate product set;

[0114] Input module 40 is used to input the target score into the regression model to obtain the comprehensive score of the candidate product;

[0115] The sorting module 50 is used to sort all candidate products in the candidate product set according to the comprehensive score, obtain a sorting result, and recommend relevant products to the target user based on the sorting result. The product recommendation device provided in this application, employing the product recommendation method in the above embodiments, can solve the technical problem of product recommendation. Compared with the prior art, the beneficial effects of the product recommendation device provided in this application are the same as those of the product recommendation method provided in the above embodiments, and other technical features in the product recommendation device are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0116] This application provides a product recommendation device, which includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, which are executed by the at least one processor to enable the at least one processor to perform the product recommendation method in Embodiment 1 above.

[0117] The following is for reference. Figure 7 The diagram illustrates a structural schematic suitable for implementing the product recommendation device in the embodiments of this application. The product recommendation device in the embodiments of this application may include, but is not limited to, mobile terminals such as mobile phones, laptops, digital broadcast receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Description), PMPs (Portable Media Players), in-vehicle terminals (e.g., in-vehicle navigation terminals), and fixed terminals such as digital TVs and desktop computers. Figure 7 The product recommendation device shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0118] like Figure 7As shown, the product recommendation device may include a processing unit 1001 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to a program stored in a read-only memory 1002 or a program loaded from a storage device 1003 into a random access memory 1004. The random access memory 1004 also stores various programs and data required for the operation of the product recommendation device. The processing unit 1001, the read-only memory 1002, and the random access memory 1004 are interconnected via a bus 1005. An input / output interface 1006 is also connected to the bus. Typically, the following systems can be connected to the input / output interface 1006: input devices 1007 including, for example, a touch screen, touchpad, keyboard, mouse, image sensor, microphone, accelerometer, gyroscope, etc.; output devices 1008 including, for example, a liquid crystal display (LCD), speaker, vibrator, etc.; storage devices 1003 including, for example, magnetic tape, hard disk, etc.; and communication devices 1009. Communication device 1009 allows the product recommendation device to communicate wirelessly or wiredly with other devices to exchange data. Although the figures show product recommendation devices with various systems, it should be understood that it is not required to implement or possess all of the systems shown. More or fewer systems may be implemented alternatively.

[0119] Specifically, according to the embodiments disclosed in this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 1003, or installed from read-only memory 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of the embodiments disclosed in this application.

[0120] The product recommendation device provided in this application, employing the product recommendation method in the above embodiments, can solve the technical problem of product recommendation. Compared with the prior art, the beneficial effects of the product recommendation device provided in this application are the same as those of the product recommendation method provided in the above embodiments, and other technical features of the product recommendation device are the same as those disclosed in the previous embodiment method, and will not be repeated here.

[0121] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0122] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0123] This application provides a computer-readable storage medium having computer-readable program instructions (i.e., a computer program) stored thereon, the computer-readable program instructions being used to execute the product recommendation method in the above embodiments.

[0124] The computer-readable storage medium provided in this application may be, for example, a USB flash drive, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems or devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system or device. The program code contained on the computer-readable storage medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.

[0125] The aforementioned computer-readable storage medium may be included in the product recommendation device; or it may exist independently and not be assembled into the product recommendation device.

[0126] The aforementioned computer-readable storage medium carries one or more programs that, when executed by a product recommendation device, cause the product recommendation device to: construct a user profile feature vector based on the target user's historical behavior data; filter a set of candidate products from a target product library based on the user profile feature vector; determine the target user's attention indicators and determine the target score for each candidate product in the candidate product set corresponding to the attention indicators; input the target score into a regression model to obtain a comprehensive score for the candidate products; sort all candidate products in the candidate product set according to the comprehensive score to obtain a sorting result, and recommend relevant products to the target user based on the sorting result.

[0127] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a Local Area Network (LAN) or a Wide Area Network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0128] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0129] The modules described in the embodiments of this application can be implemented in software or hardware. The names of the modules do not necessarily limit the functionality of the unit itself.

[0130] The readable storage medium provided in this application is a computer-readable storage medium that stores computer-readable program instructions (i.e., a computer program) for executing the above-described product recommendation method. This solves the technical problem that traditional recommendation methods suffer from insufficient accuracy due to their inability to fully capture the multi-level needs and characteristics of users. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided in this application are the same as those of the product recommendation method provided in the above embodiments, and will not be elaborated upon here.

[0131] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the product recommendation method described above.

[0132] The computer program product provided in this application can solve the technical problem that traditional recommendation methods cannot fully capture the multi-level needs and characteristics of users, resulting in insufficient accuracy of recommendation results. Compared with the prior art, the beneficial effects of the computer program product provided in this application are the same as those of the product recommendation method provided in the above embodiments, and will not be repeated here.

[0133] The above description is only a part of the embodiments of this application and does not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. A product recommendation method, characterized in that, The product recommendation method includes: Construct user profile feature vectors based on the target users' historical behavior data; Based on the user profile feature vector, a candidate product set is selected from the target product library; Determine the attention metrics of the target user, and determine the target score for the attention metrics corresponding to each candidate product in the candidate product set; The target score is input into the regression model to obtain the comprehensive score of the candidate products; All candidate products in the candidate product set are sorted according to the comprehensive score to obtain the sorting result, and relevant products are recommended to the target user based on the sorting result.

2. The product recommendation method as described in claim 1, characterized in that, The steps for constructing a user profile feature vector based on the target user's historical behavior data include: Obtain the target user's historical behavior data; Consumption characteristics, behavioral characteristics, and social characteristics are extracted from the historical behavioral data; First and second cross features are extracted from the consumption features, behavioral features, and social features; The first and second cross features are subjected to higher-order feature cross processing to generate higher-order combined features; Principal component analysis is performed on the higher-order combined features to reduce their dimensionality, resulting in low-dimensional dense features. Based on the low-dimensional dense features, the consumption features, the behavioral features, and the social features, a user profile feature vector is constructed.

3. The product recommendation method as described in claim 1, characterized in that, The step of selecting a candidate product set from the target product library based on the user profile feature vector includes: Determine the product feature vector of each product in the target product library; The user profile feature vector and the product feature vector are input into the deep matching model. The user profile feature vector is encoded by the user tower of the deep matching model to obtain the user representation vector. The product feature vector is encoded by the product tower of the deep matching model to obtain the product representation vector. Determine the multigranular similarity between the user representation vector and the product representation vector; Based on the multi-granular similarity, a set of candidate products is selected from the target product library.

4. The product recommendation method as described in claim 3, characterized in that, The step of determining the multi-granular similarity between the user representation vector and the product representation vector includes: The cosine similarity between the user representation vector and the product representation vector is used as the coarse-grained similarity. Based on the attention mechanism, the user representation vector and the product representation vector are divided into multiple sub-vector groups, and the granular similarity is determined according to the similarity between the sub-vector groups. Each sub-vector group includes user representation word vectors and product representation word vectors. Fine-grained similarity is obtained by performing element-wise interaction operations on the user representation vector and the product representation vector. Based on the coarse-grained similarity, the medium-grained similarity, and the fine-grained similarity, multi-grained similarity is determined.

5. The product recommendation method as described in claim 4, characterized in that, The step of performing element-wise interaction operations on the user representation vector with the product representation vector to obtain fine-grained similarity includes: Based on the user representation vector and the product representation vector, an element-by-element interaction vector is generated. The element-wise interaction vector, the user representation vector, and the product representation vector are concatenated to obtain an enhanced feature vector. Fine-grained similarity is obtained by performing a nonlinear transformation on the enhanced feature vector using a multilayer perceptron.

6. The product recommendation method as described in claim 1, characterized in that, The steps for determining the target user's attention metrics include: Obtain explicit and implicit feedback data from the target user; Extract the target user's specific attention indicators from the displayed feedback data; Extract the potential attention indicators of the target user from the implicit feedback data; Based on the explicitly concerned indicators and the potentially concerned indicators, an initial indicator set is generated; The timeliness weight of each indicator in the initial indicator set is determined based on the user's recent behavior data, and the importance weight of each indicator in the initial indicator set is learned through an attention mechanism. Based on the timeliness weight and the importance weight, the dynamic weight evaluation results of each indicator in the initial indicator set are determined; Based on the dynamic weight evaluation results, an optimized index set is determined according to the initial index set. The target user's focus metrics are determined from the set of optimization metrics.

7. A product recommendation device, characterized in that, The product recommendation device includes: The module is used to construct user profile feature vectors based on the target user's historical behavior data; The filtering module is used to filter out a set of candidate products from the target product library based on the user profile feature vector; The determination module is used to determine the attention indicators of the target user and to determine the target score of the attention indicators corresponding to each candidate product in the candidate product set. The input module is used to input the target score into the regression model to obtain the comprehensive score of the candidate products; The sorting module is used to sort all candidate products in the candidate product set according to the comprehensive score, obtain the sorting result, and recommend relevant products to the target user based on the sorting result.

8. A product recommendation device, characterized in that, The device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the product recommendation method as described in any one of claims 1 to 6.

9. A storage medium, characterized in that, The storage medium is a computer-readable storage medium, and a computer program is stored on the storage medium. When the computer program is executed by a processor, it implements the steps of the product recommendation method as described in any one of claims 1 to 6.

10. A computer program product, characterized in that, The computer program product includes a computer program that, when executed by a processor, implements the steps of the product recommendation method as described in any one of claims 1 to 6.

Images