A Deep Learning-Based Intelligent Recommendation System and Method for Agricultural E-commerce

By constructing a multimodal user profile and a multi-granularity graph neural network recommendation system using deep learning technology, and combining it with a lightweight deployment strategy, the system solves the problems of agricultural e-commerce recommendation systems being unable to perceive seasonal and regional characteristics and lacking high-order association mining, thus achieving accurate recommendations and cold start mitigation in various deployment environments.

CN123118216APending Publication Date: 2026-11-17NANTONG COLLEGE OF SCIENCE & TECHNOLOGY
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611416183.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-14
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing agricultural e-commerce recommendation systems cannot effectively perceive the seasonal and regional characteristics of agricultural products, lack the ability to explore the complex and high-level relationships between agricultural products and users, cannot make effective recommendations in cold start scenarios, and are difficult to adapt to various deployment environments.

Method used

It employs a deep learning-based multimodal user profile building module, a multi-granular graph neural network recommendation system module, and a lightweight deployment and edge computing module. By combining multi-dimensional quantized feature encoding, graph neural networks, and collaborative filtering models, it achieves accurate recommendations for agricultural products. The model is pruned and deployed on edge computing nodes, supporting multiple deployment environments.

Benefits of technology

It significantly improves the accuracy of recommendation results, effectively captures high-order relationships between agricultural products, alleviates the recommendation dilemma in cold start scenarios, and enables the system to operate efficiently in various deployment environments.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123118216A_ABST
    Figure CN123118216A_ABST
Patent Text Reader

Abstract

This invention relates to the field of intelligent recommendation technology in e-commerce, specifically to an intelligent recommendation system and method for agricultural e-commerce based on deep learning. The recommendation system includes a multimodal user profile construction module, a multi-granularity graph neural network recommendation system module, a lightweight deployment and edge computing module, and a closed-loop feedback and dynamic update module. This invention improves the accuracy of recommendation results by quantizing and encoding features such as the seasonality, shelf life (in days), and origin of agricultural products; it employs the SMGNN-CF model, enabling the system to effectively capture high-order relationships; it adopts a lightweight deployment strategy, pruning and quantizing the model before deploying it on edge computing nodes; and it transfers existing user profile features through KNN similarity search, generating initial item embedding vectors for new agricultural products from the aggregated features of existing agricultural products within the same category in the item category graph neural network layer, thus achieving intelligent recommendation in cold-start scenarios.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of e-commerce intelligent recommendation technology, specifically to an intelligent recommendation system and method for agricultural e-commerce based on deep learning. Background Technology

[0002] With the rapid development of society, economy and science and technology, the digital economy is profoundly reshaping the agricultural product circulation system and promoting the transformation of the agricultural industry towards digitalization and intelligence. As a core tool connecting agricultural product supply and consumer demand, the improvement of e-commerce recommendation systems plays a key role in reducing circulation losses and improving transaction efficiency. However, existing agricultural e-commerce recommendation systems may still cause invalid exposure and circulation losses due to the mismatch between recommendation results and user needs, the seasonality of agricultural products and regional characteristics.

[0003] Existing agricultural e-commerce recommendation technologies have some obvious shortcomings, such as: the non-standard characteristics of agricultural products are not included in the recommendation model; there is a lack of quantitative coding schemes for agricultural product characteristics such as seasonality, shelf life, and place of origin, which makes the model unable to perceive the seasonal and regional characteristics of agricultural products; there is a lack of mining of complex high-order relationships between agricultural products and users, making it impossible to effectively capture high-order relationships generated by seasonality and region; there is a lack of effective recommendation schemes for new products and new users in cold start scenarios; and the recommendation system is difficult to adapt to the deployment environment in rural areas, failing to take into account the practical problems of network level, terminal computing power, and power supply in rural areas. Summary of the Invention

[0004] To address the aforementioned issues, this invention provides a deep learning-based intelligent recommendation system and method for agricultural e-commerce, which solves the problems existing in the prior art, such as the inability of models to perceive the seasonal and regional characteristics of agricultural products, the lack of mining of complex high-order relationships between agricultural products and users, the inability to effectively capture high-order relationships generated by seasons and regions, the lack of effective recommendation solutions in cold start scenarios, and the failure to consider deployment in various deployment environments.

[0005] To achieve the above objectives, this invention provides a deep learning-based intelligent recommendation system for agricultural e-commerce, comprising:

[0006] Multimodal user profile construction module: used to collect basic user data, agricultural product project attribute data and user behavior data, to perform feature quantification encoding on the seasonality, shelf life days and place of origin of agricultural products in the agricultural product project attribute data, to generate multi-dimensional quantitative features of agricultural products, and to construct dynamic user profiles based on the fusion of the multi-dimensional quantitative features with basic user data and user behavior data.

[0007] The multi-granularity graph neural network recommendation system module is used to construct item relationship graphs and item category relationship graphs and integrate attention mechanisms. It uses graph neural networks to encode the item relationship graphs and item category relationship graphs in dual-path parallel encoding, and uses attention mechanisms to jointly model time, space and season. It combines gated recurrent neural networks to model user session sequences, and forms a recommendation model through deep learning. It integrates the output of graph neural networks, gated recurrent neural networks and collaborative filtering outputs to generate recommendation lists and recommendation scores. The recommendation list is then fed back to the multimodal user profile construction module to enrich dynamic user profiles.

[0008] Lightweight deployment and edge computing module: connected to the multi-granularity graph neural network recommendation system module, used to perform model pruning and system parameter quantization reduction processing on the trained recommendation model, and deploy the compressed model to edge computing nodes for inference services;

[0009] Closed-loop feedback and dynamic update module: Connected to the multimodal user profile construction module and the multigranular graph neural network recommendation system module respectively, it is used to collect real-time interaction information between users and items, update the dynamic user profile in an incremental learning manner, and perform online incremental optimization of the recommendation model.

[0010] Furthermore, the multimodal user profile building module constructs a dynamic user profile model based on the WKK architecture, which is the Word2vec-Kmeans-KNN architecture.

[0011] When building user profiles, the original user base dataset, user behavior dataset, and project attribute dataset undergo data cleaning, filtering to remove stop words, structuring, and standardization preprocessing. Then, data mining, keyword extraction, and word embedding are used to vectorize the data. The clustered and classified labels and features are added to the user profile database as supplementary information for users and projects at multiple granularities, and the user profile data is dynamically updated.

[0012] Furthermore, the multi-dimensional quantization features specifically include:

[0013] The seasonality mentioned above uses a 12-dimensional seasonal vector to measure the seasonality of agricultural products, corresponding to the 12 months of the year. The 12 months correspond to the seasonal windows of the natural growth cycle and market time of agricultural products. The matching between agricultural products and seasonality is quantified into the weight values ​​of the corresponding dimensions in the 12-dimensional seasonal vector.

[0014] The shelf life days are dynamically weighted using an exponential decay function, and the decay rate parameter of the exponential decay function is preset according to the variety of agricultural product.

[0015] The location of the place of origin is encoded into a joint embedding vector based on the regional characteristics of agricultural products, including the specific latitude and longitude of the place of origin and the climate zone information.

[0016] Furthermore, the multi-granularity neural network recommendation system module adopts a multi-granularity graph neural network combined with a collaborative filtering recommendation model, namely the SMGNN-CF model. The SMGNN-CF model includes an item graph neural network layer, an item category graph neural network layer, an attention network layer, a vector representation fusion layer, a gated recurrent neural network layer, and a collaborative filtering layer.

[0017] The project graph neural network layer is used to construct a project relationship graph with agricultural product projects as nodes and project co-occurrence relationships in user session sequences as edges, thereby capturing project associations at the user behavior level.

[0018] The project category graph neural network layer is used to construct a project category relationship graph with agricultural product projects and agricultural product categories as nodes and the affiliation relationship between projects and categories as edges, capturing the semantic association at the category affiliation level;

[0019] The collaborative filtering layer is used to obtain basic user information, find similar users based on the basic user information, and recommend interactive items of similar users to the user, in order to alleviate the cold start problem.

[0020] Furthermore, the attention network layer is used to calculate temporal attention weights, spatial attention weights, and seasonal attention weights, and to weight and fuse the temporal attention weights, spatial attention weights, and seasonal attention weights into a unified spatiotemporal attention score;

[0021] The time attention weight is calculated using an exponential decay function based on the shelf life of agricultural products in days. Agricultural products that are further from harvest time are assigned a lower time attention score.

[0022] The spatial attention weight is calculated based on the spherical distance between the user's location and the latitude and longitude of the agricultural product's origin. Agricultural products with a closer spherical distance are assigned a higher spatial attention score.

[0023] The seasonal attention weight is determined based on the matching degree between the current time and the seasonal window of agricultural products calculated from the 12-dimensional seasonal vector. The closer the current time is to the optimal market time of agricultural products, the higher the seasonal attention score is assigned.

[0024] Furthermore, the lightweight deployment and edge computing module adopts a green computing mode, specifically including the following measures:

[0025] (1) The number of system parameters is reduced. The model pruning is unstructured pruning based on the weight size, and the parameter quantization is INT8 symmetric quantization. The model parameter number and inference energy consumption are reduced by the combination of model pruning and parameter quantization.

[0026] (2) The edge computing nodes are deployed in township-level geographical areas for data processing nearby, forming a three-level collaborative inference architecture of cloud-edge computing nodes-terminal, reducing transmission energy consumption, and controlling system response latency within a preset threshold.

[0027] (3) Dynamic voltage and frequency adjustment technology is used to regulate computing power consumption, renewable energy power is given priority, and seasonal training strategy is used to schedule training tasks to the off-peak hours. After the obsolete computing equipment is professionally processed, it is used to build edge computing nodes in townships.

[0028] Furthermore, for the cold start scenario of new users, the WKK architecture of the multi-granularity graph neural network recommendation system module transfers the profile features of similar users from the existing user profiles through KNN similarity search, and the initial user profile is generated by the multimodal user profile construction module.

[0029] For the cold start scenario of new agricultural products, the multi-granularity graph neural network recommendation system module generates the initial item embedding vector of the new agricultural products by aggregating the features of existing agricultural products under the same category in the item category graph neural network layer;

[0030] This allows for the generation of initial recommendation results for new users or new agricultural products in a cold start scenario where historical interaction data is lacking.

[0031] A deep learning-based intelligent recommendation method for agricultural e-commerce, which applies the aforementioned intelligent recommendation system for agricultural e-commerce, specifically includes the following steps:

[0032] S1. Collect user basic data, agricultural product project attribute data and user behavior data, and perform feature quantification encoding on the seasonality, shelf life days and place of origin of agricultural products in the agricultural product project attribute data to generate multi-dimensional quantitative features of agricultural products.

[0033] S2. Based on the WKK architecture, multi-dimensional quantitative features and user features are adaptively weighted and fused together, and multi-dimensional quantitative features are fused with user basic data and user behavior data to build dynamic user profiles.

[0034] S3. Construct a project relationship graph and a project category relationship graph. Use a graph neural network to encode the project relationship graph and the project category relationship graph in dual parallel paths, perform node representation learning, combine a spatiotemporal attention mechanism to perform three-dimensional joint modeling of time decay factor, spatial distance and seasonal matching degree, and combine a gated recurrent neural network to model user session sequences. Collaboratively construct a recommendation model and generate a candidate recommendation set.

[0035] S4. Fuse the output of the graph neural network, the output of the gated recurrent neural network, and the output of the collaborative filtering to output a recommendation score, and generate a recommendation list based on the recommendation score;

[0036] S5. Feed the recommendation results back to the multimodal user profile construction module for incremental learning, update the user profile representation, and optimize the recommendation model online to achieve closed-loop optimization of the model;

[0037] S6. Perform model pruning and parameter quantization compression on the trained recommendation model, deploy the compressed model to edge computing nodes, process recommendation requests locally, and control the system response latency within a preset range.

[0038] The beneficial effects of this invention are:

[0039] This invention quantifies and encodes the seasonality, shelf life, and origin of agricultural products. Seasonality is encoded into a 12-dimensional seasonal vector, shelf life is dynamically weighted using an exponential decay function, and latitude and longitude of the origin and climate zone information are fused into a joint embedding vector. This enables deep learning models to quantitatively perceive the special attributes of agricultural products and effectively distinguish between in-season and out-of-season products, thereby significantly improving the accuracy of recommendation results.

[0040] This invention employs the SMGNN-CF model, which uses a dual-graph neural network coupled with a spatiotemporal attention mechanism to perform three-dimensional joint modeling of time decay factor, spatial distance, and seasonal matching degree. This enables the system to effectively capture the high-order correlation between agricultural products caused by seasonal alternation, geographical proximity, and co-occurrence of consumption scenarios.

[0041] This invention employs a lightweight deployment strategy, pruning and quantizing the model before deploying it on edge computing nodes, dynamically adjusting the voltage and frequency to increase the proportion of green electricity usage, rationally arranging training strategies, and tiered application of obsolete equipment to reduce the burden on grassroots levels and achieve the complete implementation of technical solutions from the laboratory to the field.

[0042] This invention can transfer existing user profile features through KNN similarity search, and generate initial item embedding vectors for new agricultural products by aggregating the features of existing agricultural products under the same category in the item category graph neural network layer. This enables intelligent recommendation in cold start scenarios, effectively alleviating the recommendation dilemma of new users and new agricultural products without historical data to rely on, and helps to improve the matching degree between the first recommendation results and user needs in cold start scenarios. Attached Figure Description

[0043] To more clearly illustrate the technical solution of the present invention, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0044] Figure 1 User profile building process;

[0045] Figure 2 This is a diagram of the SMGNN-CF model structure. Detailed Implementation

[0046] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0047] Example 1

[0048] The construction of the recommendation model and the recommendation method are as follows:

[0049] S1. Collect user basic data, agricultural product project attribute data and user behavior data, and perform feature quantification encoding on the seasonality, shelf life days and place of origin of agricultural products in the agricultural product project attribute data to generate multi-dimensional quantitative features of agricultural products.

[0050] Among them, seasonal coding: the seasonality of each agricultural product is coded as a 12-dimensional seasonal vector. The system determines its seasonal window based on the natural growth cycle and market time of the agricultural product. For example, the month in the best quality window is assigned a high value, the month in the market period but not in the best quality is assigned a medium value, and the month outside the seasonal window is assigned a low value.

[0051] Shelf life coding: The system takes the time when agricultural products are harvested and put into storage as the starting point, and dynamically weights the shelf life through an exponential decay function. Different types of agricultural products are assigned different decay rate parameters according to their characteristics and shelf life. The system calculates the freshness score of each agricultural product based on the current time, which is directly used as the quantitative feature of the shelf life dimension.

[0052] Origin Coding: The system encodes the location of agricultural products as a joint embedding vector that integrates the latitude and longitude of the origin with climate zone information. The latitude and longitude coordinates are converted into three-dimensional Cartesian coordinates to calculate spherical distance. At the same time, the climate zone type (such as North Subtropical, Central Subtropical, Warm Temperate, etc.) is mapped into a unique thermal coding vector. After being projected to the same dimension as the geographic location vector through a linear transformation layer, the vectors are spliced ​​and fused.

[0053] The core formula corresponding to the above encoding method is as follows:

[0054] Seasonal encoding: Season(m) = [sin(2πm / 12), cos(2πm / 12)], where m∈{1,2,…,12} represents the month of listing;

[0055] Shelf-weighted: Shelf(d) = d·e^(-0.02×t), where d is the shelf life in days and t is the number of days since harvesting. This value is zero for expired products.

[0056] Origin encoding: Origin(lat,lon) = Concat(Geohash(lat,lon), OneHot(climate)), which are mapped to 64-dimensional and 5-dimensional dimensions respectively through the Embedding layer and then concatenated.

[0057] S2, such as Figure 1 As shown, based on the WKK architecture, the system collects basic user data, user behavior data, and project attribute data. The collected data is processed, and the Word2vec model is used to map discrete features such as user identifiers, product identifiers, category tags, and user behavior into embedding vectors. The K-means algorithm is used to cluster all users, grouping users with similar purchasing behaviors and preferences into the same group. Multi-dimensional quantitative features are integrated with basic user data and user behavior data to achieve real-time perception and updating of user preferences and build dynamic user profiles that fit the user's situation.

[0058] Among them, the user static feature u_static and the behavior sequence feature h_behavior are fused through a gating mechanism:

[0059] g = σ(W_g·[u_static, h_behavior] + b_g), where σ is the Sigmoid function;

[0060] Dynamic user profile: u_dynamic = g⊙h_behavior +(1-g)⊙MLP(u_static), where ⊙ is the element-wise product;

[0061] When a new user registers, since the user's initial behavioral data is extremely limited or zero, the system uses the KNN algorithm to find the k users with the highest similarity to the user's basic information in the existing user group. The similarity calculation uses weighted Euclidean distance. The system uses the weighted average of the profile features of the k nearest neighbor users as the initial profile of the new user, which effectively alleviates the cold start problem for new users and improves the accuracy of cold start.

[0062] S3, such as Figure 2 As shown, a project relationship graph and a project category relationship graph are constructed. A graph neural network is used to encode the project relationship graph and the project category relationship graph in a dual-path parallel manner and to learn node representations. The project graph neural network layer outputs project embeddings with enhanced behavioral associations, and the project category graph neural network layer outputs project embeddings with enhanced category semantics. The new agricultural products generate initial embedding vectors by aggregating the features of existing agricultural products in the same category in the project category graph neural network layer.

[0063] By combining the spatiotemporal attention mechanism, a three-dimensional joint model of time decay factor, spatial distance and seasonal matching degree is performed, and the weights of time attention, spatial attention and seasonal attention are dynamically calculated and fused.

[0064] The vector representation fusion layer integrates the behavior association embedding, category semantic embedding, and attention weighted vector of the first three layers;

[0065] Gated recurrent neural network layers model user conversation sequences, collaboratively construct recommendation models, and generate candidate recommendation sets;

[0066] Among them, graph neural network node updates (taking the project relationship graph as an example): AGG is the mean aggregation function;

[0067] Spatiotemporal attention score (after adding a penalty term): α_vu = α_vu^base·δ_time·δ_space·δ_season, where:

[0068] Basic attention: α_vu^base =Softmax(LeakyReLU(a^T·[h_v⊕h_u]));

[0069] Time penalty: δ_time = e^(-0.01 × Δt), where Δt is the number of days between co-occurrences;

[0070] Space penalty: δ_space = clamp(1 - Haversine(loc_v, loc_u) / 1000, 0.1, 1.0);

[0071] Seasonal penalty: δ_season=(1+cos(Season(m_v)-Season(m_u))) / 2;

[0072] Session sequence modeling (GRU): h_t=GRU(h_i_t, h_t-1), taking the output of the last time step as the session representation h_GRU.

[0073] S4. The output of the graph neural network, the output of the gated recurrent neural network, and the output of the collaborative filtering are fused. The collaborative filtering layer decomposes and generates user / item collaborative filtering vectors, which are then fused with the item embeddings output by the item graph neural network layer. This enhances the collaborative filtering representation with graph structure information, outputs a recommendation score, and generates a recommendation list based on the recommendation score.

[0074] Among them, the three-way score fusion is: s_final=Sigmoid(W_2·ReLU(W_1·[s_GNN, s_GRU, s_CF]+b_1)+b_2), where s_GNN=h_GNN^T·u_dynamic, s_GRU=h_GRU^T·u_dynamic, and s_CF is the collaborative filtering output;

[0075] Training loss (BPR): L=-Σ_(u,i,j)lnσ(s_final(u,i)-s_final(u,j))+0.001||Θ||², where i is a positive sample and j is a negative sample.

[0076] The S5 module, which integrates closed-loop feedback and dynamic updates, collects real-time user interaction feedback on recommendation results (clicks, adding to cart, purchasing, ratings, reviews, etc.) and updates the dynamic user profile using incremental learning. It adjusts user profile preferences based on user interaction data. The system aggregates user interaction logs and performs online incremental optimization of the recommendation model. Based on existing model parameters, it conducts several rounds of fine-tuning training using new data to adapt to seasonal changes in user preferences and the emergence of new products. Through this closed-loop mechanism of "recommendation-feedback-update," user profiles and recommendation models become increasingly accurate over time.

[0077] Wherein, the incremental loss of EWC is: L_EWC(Θ)=L_BPR(Θ)+(γ / 2)Σ_i F_i(Θ_i -Θ_i) 2 γ=10, F_i is the diagonal element of the Fisher information matrix, and Θ_i is the parameter of the previous round.

[0078] S6. Perform model pruning and parameter quantization compression on the trained recommendation model to reduce the number of model parameters and inference energy consumption. Deploy the compressed model to edge computing nodes to form a three-level architecture of cloud-edge node-terminal. The cloud deploys the full-precision model and performs global model training and periodic updates during off-peak electricity hours. Edge computing nodes are deployed at the township level to process recommendation requests locally and control system response latency within a preset threshold. Terminal devices (mobile phones, computers, etc.) deploy a lightweight inference engine to handle recommendation result display and user interaction data collection.

[0079] The quantization formula is as follows: weight quantization: Δ is the quantization step size, z is the zero point, and FP32 weights are mapped to INT8 weights to reduce model storage size and inference resource consumption.

[0080] Example 2

[0081] When a new user logs into an agricultural e-commerce platform for the first time, the system lacks the user's historical behavior data. In this case, the cold start process based on the KNN algorithm is as follows:

[0082] (1) Construct a complete user profile database. Each profile contains the following information: user basic feature embedding vector (including basic information such as age, gender, occupation, and geographical location), user behavior feature embedding vector (including behaviors such as browsing history, purchase, collection, and rating), user preference feature vector (covering vegetable preference score, fruit preference score, aquatic product preference score, livestock and poultry preference score, grain and oil preference score, seasonal sensitivity score, and place of origin preference score), cluster group identifier (used to indicate the typical user group to which the user belongs), and recent behavior summary data (including the list of recently browsed products, the list of recently purchased products, and the keywords recently searched).

[0083] (2) Collect new user registration information, including age, gender, occupation, geographical location, preferred categories, etc.;

[0084] (3) The system converts the basic information of new users into search feature vectors and sets the weight of each search feature vector according to the characteristics of each basic information, such as geographical location weight 0.4 (agricultural products have strong regional attributes, and proximity to the place of origin is an important factor affecting consumption preferences), occupation weight 0.2 (there are systematic differences in consumption capacity and preferences among different occupational groups), age weight 0.3 (different age groups have different cooking and consumption habits), and gender weight 0.1;

[0085] (4) Based on the geographical location of the new user, perform an initial screening in the existing user database to select existing users with the same geographical location as the new user, so as to narrow the search range and reduce the amount of computation for subsequent fine calculations;

[0086] (5) Based on the existing users in the initial screening, calculate the weighted Euclidean distance between the existing users and the new users, sort them in ascending order of distance, and further select the 10 users closest to the new users as the new users' nearest neighbor group.

[0087] (6) The profile features of the above 10 nearest neighbor users are weighted and averaged to generate the initial user profile of the new user. The closer the distance, the higher the weight and the greater the contribution to the initial profile of the new user. The initial profile of the new user is stored in the user profile database and assigned a cluster group identifier to the new user.

[0088] (7) The system uses the SMGNN-CF architecture to generate a list of recommended agricultural products for new users on their first visit;

[0089] (8) After a new user completes their first visit and generates interactive behavior, the user profile of the new user is updated through closed-loop feedback and dynamic update module feedback, and incremental learning is used.

[0090] Example 3

[0091] When agricultural products that have never been sold on the platform are listed for the first time, the recommendation system lacks historical interaction data for the product. The cold start process based on the SMGNN-CF model is as follows:

[0092] (1) When a new agricultural product is listed on the platform for the first time, the project attribute data of the agricultural product is collected, including: agricultural product name, category, seasonal information (natural growth cycle, market launch window), shelf life in days, place of origin, production entity information, quality grade, product description text, etc.

[0093] (2) The system preprocesses and encodes the project attribute data of new agricultural products, encodes the seasonality into a 12-dimensional seasonal vector, assigns values ​​to the corresponding dimensions according to the market window period and corresponding quality of new agricultural products, sets the shelf life decay rate parameter through the exponential decay function, integrates the latitude and longitude of the place of origin and the climate zone information into a joint embedding vector, extracts keywords and embeds words into the name of agricultural products and product description text, generates text feature vectors, and splices the above features to form the initial project feature vector of new agricultural products;

[0094] (3) The system establishes a complete project embedding space for all existing agricultural products through the project category graph neural network layer in the SMGNN-CF model. It integrates the original features of the project, the project association information at the user behavior level, and the semantic information at the project subordination level to construct the embedding vector of each agricultural product project node. In the embedding space, agricultural products under the same category are clustered and distributed, and agricultural products that are on the market in the same season, from the same place of origin, or have similar consumption scenarios are close to each other.

[0095] (4) The system reads the category attributes of the new agricultural products, determines the category, retrieves all existing agricultural product items under the category node in the project category relationship graph, performs aggregation operation on the project embedding vectors of existing agricultural products, and calculates the average value of all existing project embedding vectors under the same category.

[0096] (5) Based on the aggregated average, further consider the representative weight of each existing item in the same category in the category, comprehensively consider the historical interaction frequency of the item (the higher the comprehensive frequency, the more popular it is) and the similarity between the item and the category center vector (the closer it is to the category semantic center, the stronger the representativeness). Through weighted average, improve the contribution of high-quality and typical items and form a more accurate category semantic center vector.

[0097] (6) The category semantic center vector and the initial item feature vector of the new agricultural product are fused together as the initial item embedding vector of the new agricultural product. The initial item embedding vector contains both the common semantic features of agricultural products of the same category and the unique attribute features of the new agricultural product itself.

[0098] (7) Based on the initial item embedding vector of the new agricultural product, the system generates a recommendation and delivery plan for the new agricultural product through the multi-granular neural network recommendation system module based on the SMGNN-CF architecture. When a user accesses the platform, the system calculates the matching degree between the user's access preferences and the initial item embedding vector of the new agricultural product, and determines whether to recommend the product to the user based on the matching degree. The system prioritizes recommending the new agricultural product to user groups with similar preferences for agricultural products, thereby achieving accurate cold start recommendation of the new agricultural product.

[0099] (8) After the new agricultural product completes its first recommendation and generates user interaction, the closed-loop feedback and dynamic update module collects interaction data and coordinates with the multi-granular neural network recommendation system module to update the item embedding vector of the agricultural product. As the interaction data accumulates, the initial item embedding vector that relies on the migration of similar items is gradually transitioned to a personalized embedding based on its own interaction data, realizing the complete evolution path from the cold start of the new agricultural product to the stable recommendation service of mature products.

[0100] In one implementation, the recommendation performance can be evaluated using both offline hold-out test sets and online A / B testing. The initial recommendation performance in a cold start scenario can be evaluated using at least one metric selected from Precision@K, Recall@K, HitRate@K, or conversion rate; model compression performance can be evaluated using parameter count, model storage volume, and inference energy consumption; deployment performance can be evaluated using end-to-end response latency. During testing, the dataset size and partitioning method, cold start object definition, K value, negative sample construction method, baseline model, hardware configuration, software version, quantization and pruning parameters, number of repeated tests, and statistical methods should be recorded. The specific results of the above metrics vary depending on the dataset, evaluation criteria, and deployment environment, and are not considered as fixed numerical limitations that must be achieved in this invention.

[0101] The preferred and optional technical means disclosed in this invention, except as otherwise specified and as further defined as one preferred or alternative technical means being another, can be arbitrarily combined to form several different technical solutions. Therefore, equivalent variations made according to the claims are still within the scope of this invention.

Claims

1. A deep learning-based intelligent recommendation system for agricultural e-commerce, characterized in that, include: Multimodal user profile construction module: used to collect basic user data, agricultural product project attribute data and user behavior data, to perform feature quantification encoding on the seasonality, shelf life days and place of origin of agricultural products in the agricultural product project attribute data, to generate multi-dimensional quantitative features of agricultural products, and to construct dynamic user profiles based on the fusion of the multi-dimensional quantitative features with basic user data and user behavior data. The multi-granularity graph neural network recommendation system module is used to construct item relationship graphs and item category relationship graphs and integrate attention mechanisms. It uses graph neural networks to encode the item relationship graphs and item category relationship graphs in dual-path parallel encoding, and uses attention mechanisms to jointly model time, space and season. It combines gated recurrent neural networks to model user session sequences, and forms a recommendation model through deep learning. It integrates the output of graph neural networks, gated recurrent neural networks and collaborative filtering outputs to generate recommendation lists and recommendation scores. The recommendation list is then fed back to the multimodal user profile construction module to enrich dynamic user profiles. Lightweight deployment and edge computing module: connected to the multi-granularity graph neural network recommendation system module, used to perform model pruning and system parameter quantization reduction processing on the trained recommendation model, and deploy the compressed model to edge computing nodes for inference services; Closed-loop feedback and dynamic update module: Connected to the multimodal user profile construction module and the multigranular graph neural network recommendation system module respectively, it is used to collect real-time interaction information between users and items, update the dynamic user profile in an incremental learning manner, and perform online incremental optimization of the recommendation model.

2. The intelligent recommendation system for agricultural e-commerce based on deep learning according to claim 1, characterized in that, The multimodal user profile building module constructs a dynamic user profile model based on the WKK architecture, which is the Word2vec-Kmeans-KNN architecture. When building user profiles, the original user base dataset, user behavior dataset, and project attribute dataset undergo data cleaning, filtering to remove stop words, structuring, and standardization preprocessing. Then, data mining, keyword extraction, and word embedding are used to vectorize the data. The clustered and classified labels and features are added to the user profile database as supplementary information for users and projects at multiple granularities, and the user profile data is dynamically updated.

3. The intelligent recommendation system for agricultural e-commerce based on deep learning according to claim 2, characterized in that, The multi-dimensional quantization features specifically include: The seasonality mentioned above uses a 12-dimensional seasonal vector to measure the seasonality of agricultural products, corresponding to the 12 months of the year. The 12 months correspond to the seasonal windows of the natural growth cycle and market time of agricultural products. The matching between agricultural products and seasonality is quantified into the weight values ​​of the corresponding dimensions in the 12-dimensional seasonal vector. The shelf life days are dynamically weighted using an exponential decay function, and the decay rate parameter of the exponential decay function is preset according to the variety of agricultural product. The location of the place of origin is encoded into a joint embedding vector based on the regional characteristics of agricultural products, including the specific latitude and longitude of the place of origin and the climate zone information.

4. The intelligent recommendation system for agricultural e-commerce based on deep learning according to claim 3, characterized in that, The multi-granularity neural network recommendation system module adopts a multi-granularity graph neural network combined with a collaborative filtering recommendation model, namely the SMGNN-CF model. The SMGNN-CF model includes an item graph neural network layer, an item category graph neural network layer, an attention network layer, a vector representation fusion layer, a gated recurrent neural network layer, and a collaborative filtering layer. The project graph neural network layer is used to construct a project relationship graph with agricultural product projects as nodes and project co-occurrence relationships in user session sequences as edges, thereby capturing project associations at the user behavior level. The project category graph neural network layer is used to construct a project category relationship graph with agricultural product projects and agricultural product categories as nodes and the affiliation relationship between projects and categories as edges, capturing the semantic association at the category affiliation level; The collaborative filtering layer is used to obtain basic user information, find similar users based on the basic user information, and recommend interactive items of similar users to the user, in order to alleviate the cold start problem.

5. The intelligent recommendation system for agricultural e-commerce based on deep learning according to claim 4, characterized in that, The attention network layer is used to calculate temporal attention weights, spatial attention weights, and seasonal attention weights, and to weight and fuse the temporal attention weights, spatial attention weights, and seasonal attention weights into a unified spatiotemporal attention score; The time attention weight is calculated using an exponential decay function based on the shelf life of agricultural products in days. Agricultural products that are further from harvest time are assigned a lower time attention score. The spatial attention weight is calculated based on the spherical distance between the user's location and the latitude and longitude of the agricultural product's origin. Agricultural products with a closer spherical distance are assigned a higher spatial attention score. The seasonal attention weight is determined based on the matching degree between the current time and the seasonal window of agricultural products calculated from the 12-dimensional seasonal vector. The closer the current time is to the optimal market time of agricultural products, the higher the seasonal attention score is assigned.

6. The intelligent recommendation system for agricultural e-commerce based on deep learning according to claim 1, characterized in that, The lightweight deployment and edge computing module adopts a green computing model, specifically including the following measures: (1) The number of system parameters is reduced. The model pruning is unstructured pruning based on the weight size, and the parameter quantization is INT8 symmetric quantization. The model parameter number and inference energy consumption are reduced by the combination of model pruning and parameter quantization. (2) The edge computing nodes are deployed in township-level geographical areas for data processing nearby, forming a three-level collaborative inference architecture of cloud-edge computing nodes-terminal, reducing transmission energy consumption, and controlling system response latency within a preset threshold. (3) Dynamic voltage and frequency adjustment technology is used to regulate computing power consumption, renewable energy power is given priority, and seasonal training strategy is used to schedule training tasks to the off-peak hours. After the obsolete computing equipment is professionally processed, it is used to build edge computing nodes in townships.

7. The intelligent recommendation system for agricultural e-commerce based on deep learning according to claim 4, characterized in that, For the cold start scenario of new users, the WKK architecture of the multi-granularity graph neural network recommendation system module transfers the profile features of similar users from the existing user profiles through KNN similarity search, and the initial user profile is generated by the multimodal user profile construction module. For the cold start scenario of new agricultural products, the multi-granularity graph neural network recommendation system module generates the initial item embedding vector of the new agricultural products by aggregating the features of existing agricultural products under the same category in the item category graph neural network layer; This allows for the generation of initial recommendation results for new users or new agricultural products in a cold start scenario where historical interaction data is lacking.

8. A deep learning-based intelligent recommendation method for agricultural e-commerce, comprising an agricultural product recommendation method applying the intelligent recommendation system for agricultural e-commerce as described in any one of claims 1 to 7, characterized in that, Specifically, the following steps are included: S1. Collect user basic data, agricultural product project attribute data and user behavior data, and perform feature quantification encoding on the seasonality, shelf life days and place of origin of agricultural products in the agricultural product project attribute data to generate multi-dimensional quantitative features of agricultural products. S2. Based on the WKK architecture, multi-dimensional quantitative features and user features are adaptively weighted and fused together, and multi-dimensional quantitative features are fused with user basic data and user behavior data to build dynamic user profiles. S3. Construct a project relationship graph and a project category relationship graph. Use a graph neural network to encode the project relationship graph and the project category relationship graph in dual parallel paths, perform node representation learning, combine a spatiotemporal attention mechanism to perform three-dimensional joint modeling of time decay factor, spatial distance and seasonal matching degree, and combine a gated recurrent neural network to model user session sequences. Collaboratively construct a recommendation model and generate a candidate recommendation set. S4. Fuse the output of the graph neural network, the output of the gated recurrent neural network, and the output of the collaborative filtering to output a recommendation score, and generate a recommendation list based on the recommendation score; S5. Feed the recommendation results back to the multimodal user profile construction module for incremental learning, update the user profile representation, and optimize the recommendation model online to achieve closed-loop optimization of the model; S6. Perform model pruning and parameter quantization compression on the trained recommendation model, deploy the compressed model to edge computing nodes, process recommendation requests locally, and control the system response latency within a preset range.