Supply chain product recommendation method and device, electronic equipment and storage medium

Abstract

The embodiment of the invention provides a supply chain product recommendation method and device, electronic equipment and a storage medium, and belongs to the technical field of supply chains. The method comprises the following steps: acquiring a target supply chain feature, a target product feature and a target warehouse distribution feature; calling an object tower to perform feature extraction on target supply chain features to obtain target supply chain embedding, calling a product tower to perform feature extraction on target product features to obtain target product embedding, and calling a warehouse distribution tower to perform feature extraction on target warehouse distribution features to obtain target warehouse distribution embedding; performing attention fusion according to target supply chain embedding, target product embedding and target warehouse matching embedding to obtain target fusion features; and calling a scoring network to score the target fusion feature to obtain a target matching score of the candidate supply chain product corresponding to the target fusion feature, and determining a target supply chain product from the plurality of candidate supply chain products according to the target matching score. According to the embodiment of the invention, the recommendation accuracy of supply chain products can be improved.

Description

Technical Field

[0001] This application relates to the field of supply chain technology, and in particular to a supply chain product recommendation method, apparatus, electronic device, and storage medium. Background Technology

[0002] Supply chain products refer to solutions that match user needs with supply chain resources (such as warehousing and distribution) to achieve optimal cost or meet delivery requirements. For example, when a user places an order for a product, next-day delivery from a specific warehouse can be recommended based on user needs and warehouse costs.

[0003] However, in related technologies, supply chain products are generally recommended manually, which can easily lead to mismatches, i.e., the recommendation accuracy is low. Summary of the Invention

[0004] The main objective of this application is to provide a supply chain product recommendation method, apparatus, electronic device, and storage medium that can improve the accuracy of supply chain product recommendations.

[0005] To achieve the above objectives, a first aspect of this application proposes a supply chain product recommendation method, the method comprising:

[0006] Obtain the target supply chain characteristics of the target object, the target product characteristics of the products to be delivered corresponding to the target object, and the target warehousing and distribution characteristics of the candidate warehouses storing the products to be delivered;

[0007] Based on the target supply chain characteristics, the target product characteristics, and the target warehousing and distribution characteristics, a pre-trained product recommendation model is invoked. The product recommendation model includes an object pyramid, a product pyramid, a warehousing and distribution pyramid, and a scoring network.

[0008] The target supply chain embedding is obtained by calling the object tower to extract features from the target supply chain features; the target product embedding is obtained by calling the product tower to extract features from the target product features; and the target warehousing and distribution embedding is obtained by calling the warehousing and distribution tower to extract features from the target warehousing and distribution features.

[0009] Attention fusion is performed based on the target supply chain embedding, the target product embedding, and the target warehousing and distribution embedding to obtain target fusion features, wherein the target fusion features are used to represent the degree of matching between the target object and the corresponding candidate supply chain products, and the candidate supply chain products include the product to be delivered and the candidate warehouse;

[0010] The scoring network is invoked to score the target fusion feature to obtain the target matching score of the candidate supply chain product corresponding to the target fusion feature. The target supply chain product is then determined from the multiple candidate supply chain products based on the target matching score.

[0011] To achieve the above objectives, a second aspect of this application provides a supply chain product recommendation device, the device comprising:

[0012] The feature acquisition unit is used to acquire the target supply chain features of the target object, the target product features of the products to be delivered corresponding to the target object, and the target warehousing and distribution features of the candidate warehouses storing the products to be delivered.

[0013] The model invocation unit is used to invoke a pre-trained product recommendation model based on the target supply chain characteristics, the target product characteristics, and the target warehousing and distribution characteristics. The product recommendation model includes an object pyramid, a product pyramid, a warehousing and distribution pyramid, and a scoring network.

[0014] The feature extraction unit is used to call the object tower to extract features from the target supply chain features to obtain the target supply chain embedding, call the product tower to extract features from the target product features to obtain the target product embedding, and call the warehousing and distribution tower to extract features from the target warehousing and distribution features to obtain the target warehousing and distribution embedding.

[0015] The feature fusion unit is used to perform attention fusion based on the target supply chain embedding, the target product embedding, and the target warehousing and distribution embedding to obtain target fusion features, wherein the target fusion features are used to represent the matching degree between the target object and the corresponding candidate supply chain products, and the candidate supply chain products include the product to be delivered and the candidate warehouse;

[0016] The product recommendation unit is used to call the scoring network to score the target fusion feature, obtain the target matching score of the candidate supply chain product corresponding to the target fusion feature, and determine the target supply chain product from multiple candidate supply chain products based on the target matching score.

[0017] To achieve the above objectives, a third aspect of the present application provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the method described in any one of the embodiments of the first aspect.

[0018] To achieve the above objectives, a fourth aspect of the present application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method described in any one of the embodiments of the first aspect.

[0019] The supply chain product recommendation method, apparatus, electronic device, and storage medium proposed in this application first integrate multi-dimensional features of the target object, the product to be delivered, and the candidate warehouse. Then, relying on a pre-trained multi-tower product recommendation model, it extracts features from each dimension separately. Next, it fuses the embedded features of the three dimensions through an attention mechanism to accurately represent the matching degree between the target object and the candidate supply chain products. Finally, it quantifies the matching score through a scoring network and selects the optimal candidate supply chain products. Thus, this application embodiment can improve the accuracy of supply chain product recommendations. Attached Figure Description

[0020] Figure 1 This is a flowchart of one embodiment of the supply chain product recommendation method provided in this application.

[0021] Figure 2 This is a flowchart of an embodiment of step S130 provided in this application;

[0022] Figure 3 This is a flowchart of an embodiment of step S140 provided in this application;

[0023] Figure 4 This is a flowchart of an embodiment of calculating the target matching score in step S150 provided in this application;

[0024] Figure 5 This is a flowchart of an embodiment of determining the target supply chain product in step S150 provided in this application;

[0025] Figure 6 This is a flowchart of an embodiment of the product recommendation model training method provided in this application.

[0026] Figure 7 This is a flowchart of an embodiment of step S660 provided in this application;

[0027] Figure 8 This is a schematic diagram of a supply chain product recommendation device provided in an embodiment of this application;

[0028] Figure 9 This is a schematic diagram of the hardware structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0029] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0030] It should be noted that although functional modules are divided in the device schematic diagram and a logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than the module division in the device or the order in the flowchart. The terms "first," "second," etc., used in the specification, claims, and the foregoing drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0031] Supply chain products refer to solutions that match user needs with supply chain resources (such as warehousing and distribution) to achieve optimal cost or meet delivery requirements. In other words, supply chain products are key factors directing the supply chain's production, transportation, and scheduling to ensure accurate product delivery based on order requirements or delivery calculations. For example, when a user places an order for a product, next-day delivery from a specific warehouse can be recommended based on user needs (such as timeliness requirements) and warehouse costs. However, current supply chain product recommendations primarily rely on human experience, leading to issues such as insufficient economic efficiency and planning bias, meaning that recommended supply chain products often have a low degree of matching with user needs.

[0032] Based on this, embodiments of this application provide a supply chain product recommendation method, apparatus, electronic device, and storage medium, which can improve the accuracy of supply chain product recommendations.

[0033] This application's embodiments can acquire and process relevant data based on artificial intelligence (AI) technology. AI is the theory, methods, technology, and application system that uses digital computers or computers-controlled machines to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results. Basic AI technologies generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing technology, operating / interactive systems, and mechatronics. AI software technologies mainly include computer vision, robotics, biometrics, speech processing, natural language processing, and machine learning / deep learning.

[0034] The supply chain product recommendation method provided in this application relates to the field of supply chain technology. This method can be applied to a terminal, a server, or software running on either a terminal or a server. In some embodiments, the terminal can be a smartphone, tablet, laptop, desktop computer, etc.; the server can be configured as an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDN), and big data and artificial intelligence platforms; the software can be an application implementing the supply chain product recommendation method, but is not limited to the above forms.

[0035] This application can be used in a wide variety of general-purpose or special-purpose computer system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network personal computers (PCs), minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices. This application can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform specific tasks or implement specific abstract data types. This application can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0036] It should be noted that in all specific embodiments of this application, when processing data related to user identity or characteristics, such as user information, user supply chain behavior data, user historical data, and user location information, user permission or consent is obtained first. Furthermore, the collection, use, and processing of this data comply with relevant laws, regulations, and standards. In addition, when embodiments of this application require access to sensitive personal information of users, separate permission or consent from the user is obtained through pop-ups or redirects to confirmation pages. Only after obtaining the user's separate permission or consent is the necessary user-related data required for the proper functioning of these embodiments acquired.

[0037] Please see Figure 1 , Figure 1This is an optional flowchart of the supply chain product recommendation method provided in the embodiments of this application. In some embodiments of this application, Figure 1 The method may include, but is not limited to, steps S110 to S150.

[0038] Step S110: Obtain the target supply chain characteristics of the target object, the target product characteristics of the products to be delivered corresponding to the target object, and the target warehousing and distribution characteristics of the candidate warehouses storing the products to be delivered.

[0039] Step S120: Based on the target supply chain characteristics, target product characteristics, and target warehousing and distribution characteristics, call the pre-trained product recommendation model. The product recommendation model includes an object pyramid, a product pyramid, a warehousing and distribution pyramid, and a scoring network.

[0040] Step S130: Call the object tower to extract features from the target supply chain features to obtain the target supply chain embedding; call the product tower to extract features from the target product features to obtain the target product embedding; call the warehousing and distribution tower to extract features from the target warehousing and distribution features to obtain the target warehousing and distribution embedding.

[0041] Step S140: Attention fusion is performed based on the target supply chain embedding, target product embedding, and target warehousing and distribution embedding to obtain target fusion features. The target fusion features are used to represent the degree of matching between the target object and the corresponding candidate supply chain products. The candidate supply chain products include products to be delivered and candidate warehouses.

[0042] Step S150: Call the scoring network to score the target fusion feature, obtain the target matching score of the candidate supply chain product corresponding to the target fusion feature, and determine the target supply chain product from multiple candidate supply chain products based on the target matching score.

[0043] In step S110 of some embodiments, the target object may refer to customers (including individuals and enterprises) who require logistics services and are seeking supply chain product recommendations. Target supply chain characteristics may refer to features describing the target object's supply chain attributes and preferences. For example, target supply chain characteristics may include customer supply chain profiles (such as the customer's industry and size), customer supply chain behavior data (such as historical order frequency and return rate), customer supply chain timeliness preferences (such as optimal timeliness, no timeliness consideration, medium timeliness with price as the most important factor, or the expected delivery time range), and customer supply chain budget preferences (such as low price, ignoring price, etc.). The product to be delivered may refer to the physical product to be transported. Target product characteristics may refer to features describing the product to be delivered. For example, target product characteristics may include basic attributes (such as product weight and volume), transportation needs (such as transportation distance and timeliness requirements), type of item to be delivered (i.e., product content, such as mobile phones and food), industry (such as the 3C industry), and scenario (such as personal parcel delivery, corporate transshipment, and large-item express delivery). Candidate warehouses refer to warehousing centers that can provide delivery services for products awaiting shipment. For example, a candidate warehouse can refer to a warehousing center that stores products awaiting shipment. Target warehousing and distribution characteristics can refer to the characteristics that describe the warehouse features and transportation conditions of candidate warehouses. For example, target warehousing and distribution characteristics can include warehouse geographical coordinates, inventory depth (such as real-time SKU inventory of different products), carrier timeliness matrix (such as historical fulfillment timeliness of different logistics service providers), inter-warehouse transfer cost matrix (such as transportation costs across regional warehouses), and anomaly factors (such as weather impacts, capacity changes during major promotions, etc.).

[0044] In step S120 of some embodiments, a pre-trained product recommendation model is invoked based on the target supply chain characteristics, target product characteristics, and target warehousing and distribution characteristics. The product recommendation model employs a multi-tower architecture, including an object tower, a product tower, and a scoring tower, and is combined with a scoring network for comprehensive evaluation. The object tower processes supply chain characteristics on the object side (e.g., the target supply chain characteristics of the target object), the product tower processes product characteristics on the product side (e.g., the target product characteristics of the product to be delivered), and the warehousing and distribution tower processes warehousing and distribution characteristics on the warehouse side (e.g., the target warehousing and distribution characteristics of the candidate warehouse). These three towers encode the features they process into fixed-dimensional vector representations for subsequent analysis. The scoring network combines the vectors output by the three towers for final matching calculations, ensuring that the recommendation results optimize supply chain costs and timeliness while meeting the object's needs.

[0045] In step S130 of some embodiments, the object tower is invoked to encode the target supply chain features to obtain the target supply chain embedding E. u =MLP u (x u ), where x uThis represents the characteristics of the target supply chain. Specifically, nonlinear encoding is performed using a multilayer perceptron to map the high-dimensional, sparse features of the object side into a low-dimensional, dense embedding vector E. u For example, E u = [0.3, -0.8, 0.67, ...], where each dimension implicitly contains information such as the target object's timeliness preference and monetary preference.

[0046] The product pyramid is invoked to encode the features of the target product, resulting in the target product embedding E. i =MLP i (x i ), where x i This represents the target product features. Specifically, similar to the object tower processing method, the target product features are mapped into low-dimensional dense embedding vectors E through independent multilayer perceptrons. i .

[0047] The distribution tower processes the target distribution features through a spatiotemporal encoder (ST-Encoder) to obtain the target distribution embedding E. l .

[0048] Reference Figure 2 In some embodiments, the step S130 of “calling the warehousing and distribution tower to extract features of the target warehousing and distribution features to obtain the target warehousing and distribution embedding” may include, but is not limited to, steps S210 to S250.

[0049] Step S210: Call the warehouse distribution tower to encode the spatial dimension of the target warehouse distribution features to obtain spatial embedding, wherein the spatial dimension includes the distance element between the address of the target object and the address of the candidate warehouse.

[0050] Step S220: Call the warehouse distribution tower to encode the target warehouse distribution features in the time dimension to obtain the time embedding, wherein the time dimension includes the historical fulfillment timeliness elements of the logistics distribution object of the product to be delivered;

[0051] Step S230: Call the warehouse distribution tower to encode the target warehouse distribution features in terms of cost dimension to obtain cost embedding, wherein the cost dimension includes the fulfillment cost element of delivering products to be delivered from the candidate warehouse;

[0052] Step S240: Call the warehouse distribution tower to encode the abnormal dimension of the target warehouse distribution features to obtain the abnormal embedding, wherein the abnormal dimension includes the warehouse distribution abnormal elements of the candidate warehouse.

[0053] Step S250: The spatial embedding, temporal embedding, cost embedding, and anomaly embedding are spliced ​​together to obtain the target warehousing and distribution embedding.

[0054] In steps S210 to S250 of this embodiment, the distribution tower can include four parallel feature extractors. That is, the distribution tower can perform domain-specific embedding of the target distribution features from four dimensions using a spatiotemporal encoder, and then concatenate the embedding results of the four dimensions to obtain the target distribution embedding. Specifically, the four dimensions are spatial dimension, temporal dimension, cost dimension, and anomaly dimension. The spatial dimension represents the distance between the address of the candidate warehouse and the query point, using the address of the target object as the query point. Thus, the spatial embedding can be determined based on this distance. For example, for each candidate warehouse, the distance between the address of the candidate warehouse and the query point can be calculated separately (e.g., calculating the Haversine distance), and the distance can be discretized and bucketed before embedding to obtain the spatial embedding corresponding to each candidate warehouse (e.g., distances from 0 to 50 km can be embedded in [0.2, 0.5]).

[0055] The time dimension can represent the SLA (Service Level Agreement) quantile of a carrier (i.e., a logistics service provider, such as the logistics delivery object in this embodiment), which reflects the carrier's timeliness performance in historical fulfillment. Specifically, the SLA quantile represents the proportion of times a carrier actually meets or exceeds the service standards agreed upon in the agreement within a statistical period. For example, when the SLA quantile is 95%, it indicates that 95% of the carrier's delivery instances meet the timeliness standards required by the agreement (such as on-time delivery rate, transportation timeliness compliance rate, etc.). Based on this indicator, WarehousePlatform maps the SLA quantile to a low-dimensional dense vector (e.g., the time embedding corresponding to 95% can be [0.8, 0.1]), thereby quantifying the carrier's timeliness reliability. The time dimension enables the model to explicitly learn the timeliness patterns in historical fulfillment data and provides interpretable feature inputs for subsequent fusion calculations.

[0056] The cost dimension can represent the fulfillment cost required to deliver goods from a candidate warehouse to its destination (such as the address of the target object) during the logistics delivery process. This fulfillment cost is mapped using methods such as threshold segmentation to obtain the corresponding cost embedding (e.g., when the fulfillment cost is < 100 yuan, the corresponding cost embedding can be [0.1, 0.4]).

[0057] Anomalies can refer to abnormal factors in warehousing and distribution that may affect the delivery time of candidate warehouses, such as abnormal weather or changes in transportation capacity during major promotions. Taking abnormal weather as an example, different weather coefficients can be set in advance for different weather conditions. For example, when a typhoon warning is issued, the corresponding anomaly embedding can be obtained based on the weather coefficient of the typhoon warning as [0.9, -1.2].

[0058] This application implements a comprehensive and semantically rich target warehousing and distribution intrusion model by encoding and fusing warehousing and distribution features in multiple dimensions, laying a key foundation for subsequent models to accurately evaluate the comprehensive warehousing and distribution capabilities of candidate warehouses.

[0059] In step S140 of some embodiments, the target supply chain embedding E is performed through a multi-head attention mechanism. u Target product embedded with E i Embedded E for target warehousing and distribution l Interactive fusion is performed to generate target fusion features R. Target fusion features R simultaneously capture the correlation between the target object's interests (i.e., the target object's needs), the product to be delivered, and the candidate warehouses. For example, if the item to be delivered is a mobile phone of brand X, the target supply chain embedding E... u The target object has a preference for "next-day delivery", and the target warehousing and distribution embedding corresponding to candidate warehouse A is E. l The feasibility and lowest cost of next-day delivery of brand X mobile phones by candidate warehouse A are demonstrated. Thus, after multi-head attention processing, the matching score of the target object, the item to be delivered (brand X mobile phone), and candidate warehouse A will be significantly enhanced under attention weights. In other words, the degree of matching between candidate supply chain products (i.e., combinations of items to be delivered + candidate warehouses, such as brand X mobile phone + candidate warehouse A) and the target object can ultimately be quantified based on the target fusion feature R. It is understandable that each candidate warehouse can be combined with the item to be delivered to form a candidate supply chain product.

[0060] Reference Figure 3 In some embodiments, step S140 may include, but is not limited to, steps S310 to S330.

[0061] Step S310: Determine the target product embedding and the target warehousing and distribution embedding as key vectors, and determine the target product embedding and the target warehousing and distribution embedding as value vectors;

[0062] Step S320: The target supply chain is embedded and determined as a query vector;

[0063] Step S330: Attention fusion is performed based on the query vector, key vector, and value vector to obtain the target fusion feature.

[0064] In steps S310 to S330 of this application embodiment, the target product is embedded in E. i Embedded E for target warehousing and distribution l The combination of these elements serves as the key vector and value vector, respectively, i.e., K = V = [E]. i E l Furthermore, embedding the target supply chain into E uThis serves as the query vector. Thus, based on the multi-head attention mechanism, the relevance weights between the query vector and the key-value vector are calculated, and the weighted aggregated value vector is generated to produce the target fusion feature R. In other words, the multi-head attention mechanism Attn(Q=E) u K = [E i E l ],V=[E i E l ]).

[0065] This application embodiment embeds the target supply chain into E u As the core of the query, the attention mechanism automatically focuses on the target supply chain embedded in E. u Strongly correlated (i.e., strongly correlated with the target's preferences) product-warehouse combinations (i.e., candidate supply chain products). Furthermore, the target warehousing and distribution is embedded with E... l (Including distance, SLA quantiles, cost, etc.) participate in the calculation through key-value vectors, ensuring that the recommended results match the actual fulfillment capabilities (e.g., filtering out candidate warehouses that are too far away). For example, when embedding E based on the target supply chain... u When determining the target audience's preference for next-day delivery, the attention mechanism assigns a target warehousing and delivery embedding E with a high SLA quantile (e.g., 95%). l Greater weighting, while filtering out candidate warehouses with excessively high fulfillment costs, enables subsequent recommendations of product-warehouse combinations that combine timeliness and cost-effectiveness based on the target fusion feature R.

[0066] In step S150 of some embodiments, for each candidate supply chain product, a scoring network is invoked to evaluate the target fusion feature R of the candidate supply chain product, generating a target matching score for the candidate supply chain product. A higher target matching score indicates a better match between the corresponding candidate supply chain product and the target object. Thus, based on the target matching score corresponding to each candidate supply chain product, a target supply chain product can be selected from multiple candidate supply chain products. The target supply chain product is then recommended to the target object.

[0067] Reference Figure 4 In some embodiments, the scoring network includes a first scoring head and a second scoring head. The step S150, "calling the scoring network to score the target fusion feature and obtaining the target matching score of the candidate supply chain product corresponding to the target fusion feature", may include, but is not limited to, steps S410 to S430.

[0068] Step S410: Obtain the candidate timeliness coefficient of the product to be delivered in the candidate warehouse, call the first scoring head to perform inner product calculation based on the target fusion feature and the candidate timeliness coefficient, and determine the predicted click probability of the target object for the candidate supply chain product corresponding to the target fusion feature based on the result of the first inner product calculation.

[0069] Step S420: Call the second scoring head to perform multi-layer perception on the target fusion features to obtain the predicted fulfillment cost of the candidate supply chain products corresponding to the target fusion features;

[0070] Step S430: Calculate the target matching score for the candidate supply chain product based on the predicted click probability, predicted fulfillment cost, and actual fulfillment cost.

[0071] In step S410 of some embodiments, the scoring network may include dual scoring heads (i.e., a first scoring head and a second scoring head). The first scoring head (e.g., a pCTR head) can predict the click-through rate (CTR) using inner product calculation and the Sigmoid function. The predicted CTR reflects the target audience's probability of interest in candidate supply chain products (i.e., the target audience's click tendency on candidate supply chain products). Specifically, for each candidate supply chain product, firstly, the candidate timeliness coefficient of the candidate warehouse in the candidate supply chain product can be obtained. The candidate timeliness coefficient reflects whether the candidate warehouse can deliver the product to be delivered within the promised timeframe. If it is within the promised timeframe, the value of the candidate timeliness coefficient can be a first preset value (e.g., 1); otherwise, the value can be a second preset value (e.g., 0). It is understood that the candidate timeliness coefficient can be determined based on historical fulfillment data of the candidate warehouse, target warehousing and distribution characteristics, etc. Thus, the candidate timeliness coefficient can be inner producted with the target fusion feature R to obtain the first inner product calculation result. Then, the result of the first inner product calculation can be mapped to the (0,1) interval using the Sigmoid function to obtain the predicted click-through rate of the target object for the candidate supply chain products. For example, the predicted click-through rate can be calculated using the following formula 1.

[0072]

[0073] Where w represents the candidate timeliness coefficient.

[0074] In step S420 of some embodiments, for the same candidate supply chain product, a second scoring head (such as a TSC-Aware head) is invoked to perform multilayer perceptron computation on the candidate supply chain product to output a predicted fulfillment cost. The predicted fulfillment cost can reflect the cost of transporting the product to be delivered from the address of the candidate warehouse to the destination. Specifically, the second scoring head can adopt a multilayer perceptron (MLP) structure. Taking a second scoring head including two fully connected layers as an example, the first layer can be a hidden layer, which introduces a nonlinear transformation through an activation function to capture the complex correlation between object preferences, product attributes, and warehousing and distribution costs in the target fusion features. The second layer can be an output layer, used to output the predicted fulfillment cost based on the processing results of the first layer. For example, the predicted performance cost can be calculated using the following formula 2.

[0075]

[0076] In step S430 of some embodiments, for each candidate supply chain product, the predicted click-through rate of that candidate supply chain product is calculated. Predicting fulfillment costs The target matching score (Score) is calculated by combining the actual fulfillment cost (C) with the target matching cost (C). The target matching score comprehensively reflects the overall attractiveness and cost feasibility of the candidate supply chain product. Specifically, as shown in Formula 3 below, the target matching score can be calculated using a multi-objective weighted fusion method.

[0077]

[0078] Wherein, α, β, and γ are hyperparameters updated online via Bayesian methods, used to dynamically balance the influence weights of predicted click-through rate, predicted fulfillment cost, and actual fulfillment cost (α, β, and γ are all greater than 0). It is a non-linear activation function used to capture the degree of deviation between the predicted performance cost and the benchmark value (set to 1 here). This is used to convert predicted fulfillment costs into a score; the lower the predicted fulfillment cost, the higher the score.

[0079] The embodiments of this application construct a candidate supply chain product matching system that takes into account both the needs of the target and the feasibility of fulfillment based on the method of determining the target matching score by predicting the click probability and predicting the fulfillment cost, providing a key decision basis for accurate recommendation.

[0080] Reference Figure 5 In some embodiments, the step S150 of "determining the target supply chain product from multiple candidate supply chain products based on the target matching score" may include, but is not limited to, steps S510 to S550.

[0081] Step S510: Sort multiple candidate supply chain products according to the target matching score, and filter the candidate supply chain products according to the sorting results and the preset screening quantity to obtain the first preliminary sorting result;

[0082] Step S520: Based on preset performance conditions, the candidate supply chain products in the first preliminary ranking result are screened to obtain the second preliminary ranking result;

[0083] Step S530: Obtain the target satisfaction of the target object with each candidate supply chain product in the second preliminary ranking result based on the historical satisfaction dataset of the target object;

[0084] Step S540: For each candidate supply chain product in the second preliminary ranking result, the ranking reward data of the candidate supply chain product is calculated based on the target satisfaction, predicted fulfillment cost and total product amount of the product to be delivered.

[0085] Step S550: Re-rank multiple candidate supply chain products in the second ranking result according to the ranking reward data to obtain the target ranking result, and determine the target supply chain product based on the target ranking result.

[0086] In step S510 of some embodiments, all candidate supply chain products are sorted in descending order based on the target matching score corresponding to each candidate supply chain product. A higher score indicates a better overall value for the corresponding candidate supply chain product in terms of target demand matching, cost, timeliness, and feasibility. Then, the top N candidate supply chain products can be selected from the sorted list according to a pre-set preset screening quantity, forming a first preliminary sorting result. This ensures that high-potential candidate supply chain products are quickly screened from a massive number of candidate supply chain products, laying the foundation for subsequent refined screening. It is understood that the specific value of the preset screening quantity N can be dynamically set according to business scenarios, etc., and is not specifically limited thereto.

[0087] In step S520 of some embodiments, the candidate supply chain products in the first preliminary ranking result are further screened using preset fulfillment conditions to exclude candidate supply chain products that cannot be fulfilled due to objective limitations. The preset fulfillment conditions can cover multi-dimensional constraints, such as inventory constraints, cost constraints, and timeliness constraints, and are not specifically limited thereto. Taking timeliness constraints as an example, it is determined whether the estimated delivery time from the candidate warehouse address to the target object address is within the target object's preferred timeliness range, and candidate supply chain products that are not within this range are filtered out. By judging each candidate supply chain product in the first preliminary ranking result through preset fulfillment conditions, filtering out candidate supply chain products that do not meet the preset fulfillment conditions, and determining the remaining candidate supply chain products as the second preliminary ranking result, it can be ensured that the subsequently recommended supply chain products have actual fulfillment capabilities.

[0088] In step S530 of some embodiments, to quantify the target object's expected satisfaction with each candidate supply chain product in the second preliminary ranking result, the target object's historical satisfaction dataset can be obtained and analyzed. Specifically, the historical satisfaction dataset can be determined based on the target object's historical interaction records (such as clicks, purchases, reviews, repeat purchases, etc.) and explicit and / or implicit satisfaction tags (such as ratings of 1 to 5 points, return rates within 7 days of order completion, customer service complaint records, etc.). For example, the target object's historical satisfaction dataset indicates that the target object has high satisfaction with next-day delivery when purchasing mobile phone products. In this way, candidate supply chain products can be associated with the target object's historical satisfaction dataset (such as associating with the satisfaction of similar products purchased in the past) to obtain a quantitative value of the target satisfaction of the candidate supply chain product. The target satisfaction value can range from 0 to 1, with a higher value indicating a higher expected satisfaction. For example, when a candidate supply chain product can achieve next-day delivery of mobile phones, the target satisfaction of the candidate supply chain product can be 0.9.

[0089] In step S540 of some embodiments, for each candidate supply chain product in the second ranking result, the target satisfaction and predicted fulfillment cost of the candidate supply chain product can be combined. The total GMV (Gross Merchandise Volume) of the products to be delivered is used to calculate ranking reward data to further optimize the recommendation order. For example, this can be used to assess target satisfaction and predict fulfillment costs. Different weights are assigned to the total product amount, and the specific values ​​of the ranking reward data are determined by weighted summation.

[0090] In step S550 of some embodiments, after determining the ranking reward data corresponding to each candidate supply chain product, the multiple candidate supply chain products in the second preliminary ranking result are re-ranked. The candidate supply chain product with the larger the ranking reward data ranks higher in the new ranking. The re-ranking process can employ a reinforcement learning Listwise re-ranker. The Listwise re-ranker is a ranking model used in information retrieval and recommendation systems. The Listwise re-ranker optimizes the candidate result list as a whole, directly learning the ranking target at the list level. In this way, the highest-ranked candidate supply chain product can be selected as the target supply chain product from the re-ranked list (i.e., the target ranking result).

[0091] The embodiments of this application realize the entire chain from high-potential candidate screening to refined ranking, improving the accuracy of matching target supply chain products with target objects.

[0092] The training method for the product recommendation model is explained below.

[0093] Reference Figure 6In some embodiments, the training method for the product recommendation model includes, but is not limited to, steps S610 to S670.

[0094] Step S610: Obtain sample data of training samples. The sample data includes sample supply chain characteristics of sample objects, sample product characteristics of sample delivery products corresponding to sample objects, and sample warehouse and distribution characteristics of sample warehouses storing sample delivery products.

[0095] Step S620: Call the object tower to extract features from the sample supply chain features to obtain the sample supply chain embedding; call the product tower to extract features from the sample product features to obtain the sample product embedding; call the warehousing and distribution tower to extract features from the sample warehousing and distribution features to obtain the sample warehousing and distribution embedding.

[0096] Step S630: Attention fusion is performed based on sample supply chain embedding, sample product embedding and sample warehousing and distribution embedding to obtain sample fusion features. The sample fusion features are used to represent the degree of matching between the sample object and the corresponding sample supply chain product. The sample supply chain product includes sample distribution product and sample warehouse.

[0097] Step S640: Obtain the sample timeliness coefficient of the product to be delivered in the sample warehouse, call the first scoring head to perform inner product calculation based on the sample fusion feature and the sample timeliness coefficient, and determine the sample click probability of the sample object for the sample supply chain product corresponding to the sample fusion feature based on the second inner product calculation result.

[0098] Step S650: Call the second scoring head to perform multi-layer perception on the sample fusion features to obtain the sample fulfillment cost of the candidate supply chain product corresponding to the sample fusion features;

[0099] Step S660: Calculate the click-through rate loss based on the sample click-through rate and actual click-through rate of the sample supply chain products; calculate the cost regression loss based on the sample fulfillment cost and actual fulfillment cost of the sample supply chain products; and calculate the representation consistency loss based on the sample supply chain embedding, sample product embedding, and sample warehousing and distribution embedding of the sample supply chain products.

[0100] Step S670: Calculate the total loss based on the representation consistency loss, cost regression loss, and click-through rate loss, and adjust the parameters of the product recommendation model based on the total loss.

[0101] The methods for obtaining sample supply chain characteristics, sample product characteristics, and sample warehousing and distribution characteristics in step S610 are similar to those for obtaining target supply chain characteristics, target product characteristics, and target warehousing and distribution characteristics in step S110. The methods for determining sample supply chain embedding, sample product embedding, and sample warehousing and distribution embedding in step S620 are similar to those for determining target supply chain embedding, target product embedding, and target warehousing and distribution embedding in step S130. The methods for determining sample fusion characteristics in step S630 are similar to those for determining target fusion characteristics in step S140. The methods for calculating sample click probability in step S640 are similar to those for calculating predicted click probability in step S410. The methods for calculating sample fulfillment costs in step S650 are similar to those for calculating predicted fulfillment costs in step S420. Therefore, steps S610 to S650 will not be described again in this embodiment.

[0102] In steps S660 to S670 of some embodiments, the prediction bias (i.e., click-through rate loss) of the product recommendation model for click-through rate can be calculated based on the sample click-through rate predicted by the product recommendation model and the actual click-through rate (i.e., the actual click-through rate) corresponding to the sample supply chain products. The prediction bias (i.e., cost loss) of the product recommendation model for fulfillment cost can be calculated based on the sample fulfillment cost predicted by the product recommendation model and the actual fulfillment cost (i.e., the actual fulfillment cost) corresponding to the sample supply chain products. The representation consistency loss is calculated based on the comparison of the three-sided embeddings of the sample data (i.e., the sample supply chain embedding corresponding to the object side, the sample product embedding corresponding to the product side, and the sample warehousing and distribution embedding corresponding to the warehouse side). Thus, by weighted fusion of multi-task losses, a total loss is generated and used to drive the optimization of the product recommendation model parameters, thereby achieving the training of the product recommendation model.

[0103] Specifically, as shown in Formula 4 below, the total loss is calculated.

[0104]

[0105] Where λ1, λ2, and λ3 are preset weight coefficients used to balance the importance of different tasks for model training. ctr y represents the actual click-through rate. log This represents the actual click-through rate. The total loss reflects the overall performance deviation of the product recommendation model in object click prediction, cost estimation, and feature alignment. The gradient of the total loss with respect to the parameters of each layer (including the three-tower structure and the scoring network) is calculated using the backpropagation algorithm. The optimizer then updates the parameters, enabling the product recommendation model to gradually reduce the total loss in subsequent iterations, thereby improving its ability to predict and recommend candidate products, and ultimately enhancing the matching degree between the final recommended supply chain products and the target user's needs.

[0106] Reference Figure 7In some embodiments, the training samples include positive samples and negative samples, and step S660 may include, but is not limited to, steps S710 to S730.

[0107] Step S710: Calculate the positive sample distance based on the sample supply chain embedding, sample product embedding, and sample warehousing and distribution embedding corresponding to the positive sample.

[0108] Step S720: Calculate the negative sample distance based on the sample supply chain embedding, sample product embedding, and sample warehousing and distribution embedding corresponding to the negative sample.

[0109] Step S730: Determine the characterization consistency loss based on the difference between the positive sample distance and the negative sample distance.

[0110] In steps S710 to S730 of some embodiments, a positive sample can refer to a sample in which there is a positive interaction between the sample object and the sample supply chain products, such as a product clicked by the sample object, or a combination of "sample delivery product + sample warehouse" that has been successfully ordered and fulfilled. A negative sample can refer to a sample in which there is a negative association between the sample object and the sample supply chain products, including ordinary negative samples where the sample object has not generated positive behavior (such as not clicking or not purchasing), and hard negative samples where "logistics is unreachable" (such as the sample warehouse being too far from the sample object's address to deliver). Specifically, the sample supply chain embedded in the positive sample is E. u Sample product embedding and sample warehousing embedding It can be extracted from historical positive feedback data. The sample supply chain embedding E corresponding to the negative sample. u Sample product embedding and sample warehousing embedding These can be extracted from historical negative feedback data or determined through random sampling. For example, for ordinary negative samples, it can be a combination obtained through random sampling (such as random sample products + random warehouses). For hard negative samples, a combination of sample products that are the same as those of positive samples and sample warehouses that are different from those of positive samples and are not logistically accessible can be selected.

[0111] Thus, as shown in Equation 5, the distance between the embedding vectors of positive and negative samples can be calculated separately, and the final representation consistency loss can be determined by the positive sample distance and the negative sample distance.

[0112]

[0113] Where m is a preset positive parameter (with a value of 0.1 or 1), which is used to force the distance difference between positive and negative samples to be at least m, thereby reducing the insufficient discrimination learned by the model.

[0114] In this embodiment, the calculation of positive sample distance captures the association between the sample object and the effective "product + warehouse" combination, while negative sample distance reflects the characteristic differences of invalid "product + warehouse" combinations. Thus, the difference between positive and negative sample distances constrains the consistency loss, enabling the product recommendation model to learn a more discriminative embedding space. This process enhances the product recommendation model's ability to understand the true needs of the object (e.g., identifying combinations where the object prefers high-timeliness) and its ability to filter invalid candidate combinations (e.g., excluding combinations where logistics are inaccessible), ultimately improving the accuracy of the product recommendation model.

[0115] Furthermore, during the model training phase, an online distillation strategy can be employed, where a large model (the teacher model) guides the training of a lightweight three-tower model (the student model). This effectively reduces the size of the product recommendation model while ensuring accuracy loss is less than 1%, and improves inference efficiency in practical applications.

[0116] The supply chain product recommendation method provided in this application relies on multi-source data of the object (including customer behavior, product attributes, warehousing and distribution dynamics, etc.) to extract object-level product features through automated feature engineering. This achieves intelligent fusion of object and product dual-domain features, generating precise customer-product profiles tailored to each individual customer. Furthermore, by introducing a three-tower interaction architecture and a multi-head attention mechanism, the dynamic constraints of real-time supply chain inventory, timeliness, and cost are transformed into learnable high-dimensional structured vectors, effectively solving the problem of insufficient modeling of complex supply chain constraints in related technologies. Through deep fusion and joint optimization of supply chain product ranking strategies using multi-dimensional features, the method can directly output optimal supply chain product recommendations that meet both customer needs and preferences (such as timeliness sensitivity) and efficient fulfillment requirements, ultimately significantly improving the accuracy and business adaptability of the recommendation results.

[0117] Reference Figure 8 This application also provides a supply chain product recommendation device, the supply chain product recommendation device 800 including:

[0118] The feature acquisition unit 810 is used to acquire the target supply chain features of the target object, the target product features of the products to be delivered corresponding to the fulfillment cost target object, and the target warehousing and distribution features of the candidate warehouses storing the fulfillment cost products to be delivered.

[0119] Model invocation unit 820 is used to invoke a pre-trained product recommendation model based on the supply chain characteristics, product characteristics, and warehousing and distribution characteristics of the fulfillment cost target. The fulfillment cost product recommendation model includes an object tower, a product tower, a warehousing and distribution tower, and a scoring network.

[0120] The feature extraction unit 830 is used to call the fulfillment cost object tower to extract features from the fulfillment cost target supply chain features to obtain the target supply chain embedding, call the fulfillment cost product tower to extract features from the fulfillment cost target product features to obtain the target product embedding, and call the fulfillment cost warehousing and distribution tower to extract features from the fulfillment cost target warehousing and distribution features to obtain the target warehousing and distribution embedding.

[0121] The feature fusion unit 840 is used to perform attention fusion based on the fulfillment cost target supply chain embedding, fulfillment cost target product embedding, and fulfillment cost target warehousing and distribution embedding to obtain target fusion features. The fulfillment cost target fusion features are used to represent the degree of matching between the target object and the corresponding candidate supply chain products. The fulfillment cost candidate supply chain products include fulfillment cost products to be delivered and fulfillment cost candidate warehouses.

[0122] Product recommendation unit 850 is used to call the fulfillment cost scoring network to score the fulfillment cost target fusion features, obtain the target matching score of the fulfillment cost candidate supply chain products corresponding to the fulfillment cost target fusion features, and determine the target supply chain product from multiple fulfillment cost candidate supply chain products based on the fulfillment cost target matching score.

[0123] It should be noted that the supply chain product recommendation device provided in this application embodiment is used to implement the supply chain product recommendation method provided in the above embodiment, and the specific implementation process corresponds to the supply chain product recommendation method in the above embodiment. It can be referred to the aforementioned supply chain product recommendation method, and will not be repeated here.

[0124] This application also provides an electronic device (i.e., a computer device), which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement any of the supply chain product recommendation methods described in the above embodiments. This electronic device can be any smart terminal, including tablet computers, in-vehicle computers, etc.

[0125] Please see Figure 9 , Figure 9 This illustration shows the hardware structure of an electronic device according to another embodiment, the electronic device comprising:

[0126] The processor 910 can be implemented using a general-purpose central processing unit (CPU), microprocessor, application specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this application.

[0127] The memory 920 can be implemented as a read-only memory (ROM), static storage device, dynamic storage device, or random access memory (RAM). The memory 920 can store the operating system and other applications. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory 920 and called and executed by the processor 910 using the supply chain product recommendation method of the embodiments of this application.

[0128] The input / output interface 930 is used to implement information input and output;

[0129] The communication interface 940 is used to enable communication and interaction between this device and other devices. Communication can be achieved through wired means (such as USB, network cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.).

[0130] Bus 950 transmits information between various components of the device (e.g., processor 910, memory 920, input / output interface 930, and communication interface 940);

[0131] The processor 910, memory 920, input / output interface 930 and communication interface 940 are connected to each other within the device via bus 950.

[0132] This application also provides a computer-readable storage medium storing a computer program for causing a computer to execute the supply chain product recommendation method described in the above embodiments.

[0133] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs and non-transitory computer-executable programs. Furthermore, memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, memory may optionally include memory remotely located relative to the processor, and these remote memories can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0134] This invention also provides a computer program product that stores program instructions, which, when executed by a computer, cause the computer to implement the supply chain product recommendation method described in any of the above embodiments.

[0135] The embodiments described in this application are for the purpose of more clearly illustrating the technical solutions of the embodiments of this application, and do not constitute a limitation on the technical solutions provided by the embodiments of this application. As those skilled in the art will know, with the evolution of technology and the emergence of new application scenarios, the technical solutions provided by the embodiments of this application are also applicable to similar technical problems.

[0136] Those skilled in the art will understand that the technical solutions shown in the figures do not constitute a limitation on the embodiments of this application, and may include more or fewer steps than shown, or combine certain steps, or different steps.

[0137] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0138] Those skilled in the art will understand that all or some of the steps in the methods disclosed above, as well as the functional modules / units in the systems and devices, can be implemented as software, firmware, hardware, or suitable combinations thereof.

[0139] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0140] It should be understood that in this application, "at least one (item)" means one or more, and "more than" means two or more. "And / or" is used to describe the relationship between related objects, indicating that three relationships can exist. For example, "A and / or B" can represent three cases: only A exists, only B exists, and both A and B exist simultaneously, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one (item) of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one (item) of a, b, or c can represent: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.

[0141] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of the units described above is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0142] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0143] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0144] If the integrated unit is implemented as 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 this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes multiple instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing programs, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0145] The preferred embodiments of the present application have been described above with reference to the accompanying drawings, but this does not limit the scope of the claims of the present application. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and substance of the embodiments of the present application shall be within the scope of the claims of the present application.

Claims

1. A supply chain product recommendation method, characterized in that, The method includes: Obtain the target supply chain characteristics of the target object, the target product characteristics of the products to be delivered corresponding to the target object, and the target warehousing and distribution characteristics of the candidate warehouses storing the products to be delivered; Based on the target supply chain characteristics, the target product characteristics, and the target warehousing and distribution characteristics, a pre-trained product recommendation model is invoked. The product recommendation model includes an object pyramid, a product pyramid, a warehousing and distribution pyramid, and a scoring network. The target supply chain embedding is obtained by calling the object tower to extract features from the target supply chain features; the target product embedding is obtained by calling the product tower to extract features from the target product features; and the target warehousing and distribution embedding is obtained by calling the warehousing and distribution tower to extract features from the target warehousing and distribution features. Attention fusion is performed based on the target supply chain embedding, the target product embedding, and the target warehousing and distribution embedding to obtain target fusion features, wherein the target fusion features are used to represent the degree of matching between the target object and the corresponding candidate supply chain products, and the candidate supply chain products include the product to be delivered and the candidate warehouse; The scoring network is invoked to score the target fusion feature to obtain the target matching score of the candidate supply chain product corresponding to the target fusion feature. The target supply chain product is then determined from the multiple candidate supply chain products based on the target matching score.

2. The method according to claim 1, characterized in that, The scoring network includes a first scoring head and a second scoring head. Calling the scoring network to score the target fusion feature and obtaining the target matching score of the candidate supply chain product corresponding to the target fusion feature includes: Obtain the candidate timeliness coefficient of the candidate warehouse for the product to be delivered, call the first scoring head to perform inner product calculation based on the target fusion feature and the candidate timeliness coefficient, and determine the predicted click probability of the target object for the candidate supply chain product corresponding to the target fusion feature based on the first inner product calculation result; The second scoring head is invoked to perform multi-layer perception on the target fusion feature to obtain the predicted fulfillment cost of the candidate supply chain product corresponding to the target fusion feature; The target matching score for the candidate supply chain product is calculated based on the predicted click probability, predicted fulfillment cost, and actual fulfillment cost.

3. The method according to claim 2, characterized in that, The step of determining the target supply chain product from the multiple candidate supply chain products based on the target matching score includes: The candidate supply chain products are sorted according to the target matching score, and the candidate supply chain products are filtered according to the sorting result and the preset screening quantity to obtain a first preliminary sorting result. The candidate supply chain products in the first preliminary ranking result are screened according to the preset performance conditions to obtain the second preliminary ranking result; Based on the historical satisfaction dataset of the target object, obtain the target satisfaction of the target object for each candidate supply chain product in the second preliminary ranking result; For each candidate supply chain product in the second preliminary ranking result, the ranking reward data of the candidate supply chain product is calculated based on the target satisfaction, the predicted fulfillment cost and the total product amount of the products to be delivered. Based on the ranking reward data, multiple candidate supply chain products in the second ranking result are re-ranked to obtain the target ranking result, and the target supply chain product is determined based on the target ranking result.

4. The method according to claim 1, characterized in that, The step of calling the warehousing and distribution tower to extract features from the target warehousing and distribution features to obtain the target warehousing and distribution embedding includes: The warehouse distribution tower is invoked to encode the spatial dimension of the target warehouse distribution features to obtain spatial embedding, wherein the spatial dimension includes the distance element between the address of the target object and the address of the candidate warehouse; The warehouse distribution tower is invoked to encode the target warehouse distribution features in the time dimension to obtain a time embedding, wherein the time dimension includes the historical fulfillment timeliness elements of the logistics distribution object of the product to be delivered; The warehousing and distribution tower is invoked to encode the target warehousing and distribution features in the cost dimension to obtain cost embedding, wherein the cost dimension includes the fulfillment cost element of delivering the product to be delivered from the candidate warehouse; The warehouse distribution tower is invoked to encode the anomaly dimension of the target warehouse distribution features to obtain an anomaly embedding, wherein the anomaly dimension includes the warehouse distribution anomaly elements of the candidate warehouse; The spatial embedding, the temporal embedding, the cost embedding, and the anomaly embedding are concatenated to obtain the target warehousing and distribution embedding.

5. The method according to claim 1, characterized in that, The attention fusion based on the target supply chain embedding, the target product embedding, and the target warehousing and distribution embedding to obtain target fusion features includes: The target product embedding and the target warehousing and distribution embedding are determined as key vectors, and the target product embedding and the target warehousing and distribution embedding are determined as value vectors; The target supply chain is embedded and determined as a query vector; Attention fusion is performed based on the query vector, the key vector, and the value vector to obtain the target fusion feature.

6. The method according to claim 2, characterized in that, The training method for the product recommendation model includes: Obtain sample data for training samples, including sample supply chain characteristics of sample objects, sample product characteristics of sample delivery products corresponding to the sample objects, and sample warehouse and distribution characteristics of sample warehouses storing the sample delivery products. The object tower is invoked to extract features from the sample supply chain features to obtain the sample supply chain embedding; the product tower is invoked to extract features from the sample product features to obtain the sample product embedding; and the warehousing and distribution tower is invoked to extract features from the sample warehousing and distribution features to obtain the sample warehousing and distribution embedding. Attention fusion is performed based on the sample supply chain embedding, the sample product embedding, and the sample warehousing and distribution embedding to obtain sample fusion features, wherein the sample fusion features are used to represent the degree of matching between the sample object and the corresponding sample supply chain product, and the sample supply chain product includes the sample distribution product and the sample warehouse; Obtain the sample timeliness coefficient of the sample warehouse for the product to be delivered, call the first scoring head to perform inner product calculation based on the sample fusion feature and the sample timeliness coefficient, and determine the sample click probability of the sample object for the sample supply chain product corresponding to the sample fusion feature based on the second inner product calculation result; The second scoring head is invoked to perform multi-layer perception on the sample fusion features to obtain the sample fulfillment cost of the candidate supply chain product corresponding to the sample fusion features; Click-through rate loss is calculated based on the sample click-through rate and actual click-through rate of the sample supply chain products. Cost regression loss is calculated based on the sample fulfillment cost and actual fulfillment cost of the sample supply chain products. Characteristic consistency loss is calculated based on the sample supply chain embedding, sample product embedding, and sample warehousing and distribution embedding of the sample supply chain products. The total loss is calculated based on the representation consistency loss, the cost regression loss, and the click-through rate loss, and the parameters of the product recommendation model are adjusted based on the total loss.

7. The method according to claim 6, characterized in that, The training samples include positive and negative samples. The consistency loss is calculated based on the sample supply chain embedding, sample product embedding, and sample warehousing and distribution embedding of the sample supply chain products, including: The positive sample distance is calculated based on the sample supply chain embedding, sample product embedding, and sample warehousing and distribution embedding corresponding to the positive sample. The negative sample distance is calculated based on the sample supply chain embedding, sample product embedding, and sample warehousing and distribution embedding corresponding to the negative sample. The representation consistency loss is determined based on the difference between the positive sample distance and the negative sample distance.

8. A supply chain product recommendation device, characterized in that, The device includes: The feature acquisition unit is used to acquire the target supply chain features of the target object, the target product features of the products to be delivered corresponding to the target object, and the target warehousing and distribution features of the candidate warehouses storing the products to be delivered. The model invocation unit is used to invoke a pre-trained product recommendation model based on the target supply chain characteristics, the target product characteristics, and the target warehousing and distribution characteristics. The product recommendation model includes an object pyramid, a product pyramid, a warehousing and distribution pyramid, and a scoring network. The feature extraction unit is used to call the object tower to extract features from the target supply chain features to obtain the target supply chain embedding, call the product tower to extract features from the target product features to obtain the target product embedding, and call the warehousing and distribution tower to extract features from the target warehousing and distribution features to obtain the target warehousing and distribution embedding. The feature fusion unit is used to perform attention fusion based on the target supply chain embedding, the target product embedding, and the target warehousing and distribution embedding to obtain target fusion features, wherein the target fusion features are used to represent the matching degree between the target object and the corresponding candidate supply chain products, and the candidate supply chain products include the product to be delivered and the candidate warehouse; The product recommendation unit is used to call the scoring network to score the target fusion feature, obtain the target matching score of the candidate supply chain product corresponding to the target fusion feature, and determine the target supply chain product from multiple candidate supply chain products based on the target matching score.

9. An electronic device, characterized in that, The electronic device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the method according to any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 7.

Images