Intelligent bidding and tendering information pushing system and method based on RNN and WideDeep
The intelligent bidding information push system based on RNN and Wide&Deep solves the problems of low efficiency, insufficient accuracy and timeliness in enterprises' acquisition of bidding information, and realizes efficient and accurate information push and decision support, thereby improving the success rate of bidding.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-04
- Publication Date
- 2026-04-10
AI Technical Summary
In existing technologies, enterprises face problems such as low information screening efficiency, information asymmetry, insufficient accuracy in demand matching, untimely information acquisition, and lack of prediction and decision support when obtaining bidding information, making it difficult to achieve intelligent analysis and timely and accurate push notifications.
An intelligent bidding information push system based on RNN and Wide&Deep is adopted. Through data collection and preprocessing, user profile construction, bidding information feature extraction, intelligent matching and recommendation, and information push module, combined with the Long Short-Term Memory Network Model LSTM and the deep learning recommendation model Wide&Deep, the system realizes the automated processing and personalized recommendation of bidding information.
It enables efficient and accurate extraction and matching of relevant information from massive amounts of data, improving information acquisition efficiency and matching accuracy. It provides analytical support for project competitive landscape and bidding probability, ensuring the timeliness and comprehensiveness of information acquisition, reducing the risk of missing business opportunities, and increasing the bidding success rate.
Smart Images
Figure CN121834045A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of information processing technology, specifically to an intelligent bidding information push system and method based on RNN and Wide&Deep. Background Technology
[0002] In the current economic environment, participating in various bidding activities is a crucial way for enterprises to acquire projects and expand their business. However, the amount of online bidding information is vast and updates rapidly, and traditional methods of obtaining this information mainly rely on manually browsing announcements published on various public bidding platforms, government websites, or industry associations. This method has the following significant drawbacks: 1. Inefficient information filtering: The amount of bidding information is huge and the sources are scattered. Users need to spend a lot of time manually searching and filtering from multiple platforms, which is inefficient and easy to miss important information.
[0003] 2. Information asymmetry: Users have difficulty fully grasping all potential bidding opportunities that are suitable for them, especially information released in some emerging fields or specific regions.
[0004] 3. Insufficient accuracy in demand matching: Existing systems mostly match based on keywords or simple rules, lacking in-depth mining and analysis of multi-dimensional information such as user history, business preferences, qualifications and capabilities, making it difficult to achieve accurate push notifications.
[0005] 4. Untimely information access: Bidding information is highly time-sensitive, and traditional methods make it difficult to capture the latest announcements in a timely manner, causing users to miss bidding opportunities.
[0006] 5. Lack of prediction and decision support: Existing systems generally lack the ability to intelligently analyze and predict aspects such as project competition, probability of winning bids, and potential risks, and cannot provide users with effective decision-making references.
[0007] Therefore, there is an urgent need for a technical solution that can intelligently analyze user needs and accurately extract, match, and promptly push relevant information from massive amounts of bidding information. Summary of the Invention
[0008] One of the objectives of this invention is to provide an intelligent bidding information push system based on RNN and Wide&Deep, which can intelligently analyze users' bidding needs and accurately extract, match and push relevant bidding information from massive amounts of bidding information in a timely manner.
[0009] To achieve the above objectives, an intelligent bidding information push system based on RNN and Wide&Deep is provided, including: Data acquisition and preprocessing module: used to collect bidding information and preprocess the collected bidding information; User profile building module: used to acquire user data and build multi-dimensional user profiles based on the user data; The bidding information feature extraction module is used to extract feature vectors from preprocessed bidding texts using the Long Short-Term Memory (LSTM) network model based on RNN to obtain text semantic vectors; it is also used to extract non-textual structured features from bidding texts; and it is also used to concatenate the text semantic vectors output by the LSTM network model with the non-textual structured features to form the key feature vectors of the project. Intelligent matching and recommendation module: used for matching and recommending based on user profiles and key feature vectors of projects; the matching and recommendation uses the deep learning-based recommendation model Wide&Deep, taking the key feature vectors of projects and user profiles as inputs to the recommendation model Wide&Deep, to generate a personalized list of bidding information recommendations; Information push module: used to push a list of recommended bidding information to users according to a preset push strategy; Feedback and Model Optimization Module: This module collects user feedback data on the recommended list of bidding information and uses the feedback data to iteratively update and optimize the Long Short-Term Memory (LSTM) network model and the recommendation model Wide&Deep.
[0010] Furthermore, the data acquisition and preprocessing module automatically collects bidding information through various channels, including major bidding platforms, government announcements, and industry websites. The collected bidding information includes, but is not limited to, project name, issuing agency, release time, deadline, project summary, technical requirements, budget amount, and geographical location. The preprocessing includes cleaning, deduplication, word segmentation, part-of-speech tagging, and removal of stop words from the text data.
[0011] Furthermore, in the user profile building module, the user data includes the user's registration information, browsing history, search records, click behavior, winning project records, enterprise qualifications, industry preferences, and geographical location preferences; the user profile includes the user's industry, technology field, project type, budget range, geographical region, and institutions they have cooperated with.
[0012] Furthermore, the bidding information feature extraction module includes the following sub-modules: The deep semantic understanding submodule utilizes the Long Short-Term Memory (LSTM) network model to perform sequence modeling on the preprocessed bidding text. Specifically, the Word2Vec word vector model is first used to convert the words in the bidding text into low-dimensional dense vector representations, which serve as the input to the LSTM network model. Non-text structured feature extraction submodule: used to extract non-text structured features, including time features, numerical features, and category features; time features: release time, deadline, and can further calculate derived time features such as project cycle and remaining time; numerical features: budget amount, and normalization processing; category features: publishing organization, geographical location, and project type; Feature fusion submodule: concatenates text semantic vectors with non-text structured features.
[0013] Furthermore, the recommendation model Wide&Deep includes a Wide part and a Deep part; The Wide part is used to form new features through cross-feature transformation of user features and bidding information features; The Deep part contains a deep neural network (DNN) consisting of multiple fully connected layers. Its input layer is used to receive and concatenate text semantic vectors from the text semantic understanding module and non-text structured features processed by the non-text structured feature extraction module. The recommendation model Wide&Deep fuses the outputs of the Wide and Deep parts, and outputs the user's preference probability for bidding information through the Sigmoid activation function; then, it sorts the preference probabilities of bidding information to obtain a recommendation list of bidding information.
[0014] Furthermore, the information push module's push channels include in-system messages, emails, SMS messages, and App push notifications.
[0015] The second objective of this invention is to provide an intelligent bidding information push method based on RNN and Wide&Deep, comprising the following steps: Data collection and preprocessing steps: Collect bidding information and preprocess the collected bidding information; User profile building steps: Obtain user data and build multi-dimensional user profiles based on the user data; The steps for extracting features from bidding information are as follows: The preprocessed bidding text is extracted using the Long Short-Term Memory (LSTM) network model based on RNN to obtain a text semantic vector; it is also used to extract non-textual structured features from the bidding text; and it is also used to concatenate the text semantic vector output by the LSTM network model with the non-textual structured features to form the key feature vector of the project. Intelligent matching and recommendation steps: Matching and recommendation are performed based on user profiles and key feature vectors of projects; the matching and recommendation uses the deep learning-based recommendation model Wide&Deep, which takes key feature vectors of projects and user profiles as input to the recommendation model Wide&Deep to generate a personalized list of bidding information recommendations; Information push steps: Push the list of recommended bidding information to users according to the preset push strategy; Feedback and model optimization steps: Collect user feedback data on the bidding information recommendation list, and use the feedback data to iteratively update and optimize the Long Short-Term Memory (LSTM) network model and the recommendation model Wide&Deep.
[0016] Furthermore, the bidding information feature extraction module includes the following sub-modules: The deep semantic understanding submodule utilizes the Long Short-Term Memory (LSTM) network model to perform sequence modeling on the preprocessed bidding text. Specifically, the Word2Vec word vector model is first used to convert the words in the bidding text into low-dimensional dense vector representations, which serve as the input to the LSTM network model. Non-text structured feature extraction submodule: used to extract non-text structured features, including time features, numerical features, and category features; time features: release time, deadline, and can further calculate derived time features such as project cycle and remaining time; numerical features: budget amount, and normalization processing; category features: publishing organization, geographical location, and project type; Feature fusion submodule: concatenates text semantic vectors with non-text structured features.
[0017] Furthermore, the recommendation model Wide&Deep includes a Wide part and a Deep part; The Wide part is used to form new features through cross-feature transformation of user features and bidding information features; The Deep part contains a deep neural network (DNN) consisting of multiple fully connected layers. Its input layer is used to receive and concatenate text semantic vectors from the text semantic understanding module and non-text structured features processed by the non-text structured feature extraction module. The recommendation model Wide&Deep fuses the outputs of the Wide and Deep parts, and outputs the user's preference probability for bidding information through the Sigmoid activation function; then, it sorts the preference probabilities of bidding information to obtain a recommendation list of bidding information.
[0018] Principles and advantages: 1. This solution replaces the traditional method of manual browsing and filtering by automating data collection, cleaning and structured preprocessing. It can process massive amounts of multi-source bidding information in real time. Users can quickly obtain bidding information that highly matches their needs without having to search for a needle in a haystack, which greatly reduces users' time and energy costs and improves information acquisition efficiency.
[0019] 2. By using LSTM networks to perform deep semantic understanding of bidding documents and combining Wide&Deep models to collaboratively learn user profiles and project features, we can capture the implicit relationship between complex user preferences and projects, thereby providing personalized recommendations with significantly higher accuracy than traditional keyword matching, effectively improving information matching degree, and thus achieving high-precision personalized matching.
[0020] 3. This solution is designed with a closed-loop feedback learning mechanism, which can continuously iterate and optimize the recommendation model and feature extraction model based on explicit and implicit feedback from users regarding the recommendation results, such as clicks, registrations, and successful bids. This enables the system to have self-evolution capabilities and continuously adapt to changes in user needs and market dynamics, providing more timely and relevant information.
[0021] 4. Provides multi-dimensional analysis and decision support: Through semantic analysis and fusion of structured features of bidding documents, the system can not only perform information matching, but also provide potential data analysis support for project competitive situation, user winning probability, etc., to assist users in making bidding decisions, which goes beyond the function of traditional systems that only provide information listing.
[0022] 5. Ensuring Timely and Comprehensive Information Acquisition: Automated collection and real-time processing mechanisms ensure that bidding information is quickly captured and incorporated into the recommendation process. Combined with personalized push strategies, highly relevant information is proactively and promptly delivered to users, reducing the risk of missing business opportunities due to information delays or omissions. This allows users to access potential business opportunities earlier and more accurately, thereby increasing the likelihood of winning bids and improving the success rate. It also helps users discover new projects that might be overlooked by traditional methods but align with their potential capabilities and development direction. Attached Figure Description
[0023] Figure 1 This is a logical block diagram of an intelligent bidding information push system based on RNN and Wide&Deep, according to an embodiment of the present invention. Figure 2 This is a flowchart of the intelligent bidding information push method based on RNN and Wide&Deep according to an embodiment of the present invention. Detailed Implementation
[0024] The following detailed description illustrates the specific implementation method: Example A smart bidding information push system based on RNN and Wide&Deep, basically as follows: Figure 1 As shown, it includes: Data acquisition and preprocessing module: used to collect bidding information and preprocess the collected bidding information; the data acquisition and preprocessing module specifically includes a data acquisition module ( Figure 1 (B) and preprocessing module ( Figure 1 (C in the text); The data acquisition module automatically collects bidding information, and its data source / channel ( Figure 1 A) in this context includes major bidding platforms, government announcements, and industry websites. The collected bidding information includes, but is not limited to, project name, issuing agency, publication time, deadline, project summary, technical requirements, budget amount, and geographical location. The preprocessing module's preprocessing operations include cleaning the text data (removing HTML tags, special characters, advertising information, etc.), deduplication, word segmentation, part-of-speech tagging, and stop word removal to prepare for subsequent NLP processing. The preprocessed data is stored in the data storage and management module (…). Figure 1 In D) of the middle.
[0025] User profile building module ( Figure 1 F in the module is used to acquire user data and construct multi-dimensional user profiles based on that data. The user profile construction module includes user registration information, browsing history (projects viewed), search records (keywords searched), click behavior (clicked links, saved projects), bidding records (projects actively bid on and their final winning status), company qualifications, industry preferences, and geographic location preferences. The user profile includes the user's areas of interest: industry, technology field, project type, budget range, geographic region, and institutions they have collaborated with. The resulting multi-dimensional user profile is stored in a feature vector database (…). Figure 1 Part of H in (the text).
[0026] Bidding Information Feature Extraction Module Figure 1 G in the text refers to the feature vector extraction process using a Long Short-Term Memory (LSTM) network model based on an RNN. This process extracts the semantic vectors from the preprocessed bidding text and stores them in a feature vector database. Figure 1 The H in the text is also used for extracting non-textual structured features from bidding documents; it is also used to concatenate the text semantic vector output by the Long Short-Term Memory (LSTM) network model with the non-textual structured features to form the key feature vector of the project; the bidding information feature extraction module includes the following sub-modules: The deep semantic understanding submodule employs a recurrent neural network (RNN) model. This invention uses a long short-term memory (LSTM) network to perform sequence modeling on the preprocessed bidding text (project description, technical requirements, etc.). First, the Word2Vec word vector model is used to convert the words in the bidding text into low-dimensional dense vector representations, which serve as the input to the LSTM network model. The LSTM can effectively capture long-term dependencies and contextual information in the text, and output vector representations that can represent the deep semantics of the text.
[0027] (1) First, the Word2Vec word vector model is used for pre-training to obtain the word vector matrix. Where V is the vocabulary size and d is the word vector dimension. The preprocessed word sequence... Convert to word vector sequence ,in , as input to the LSTM model.
[0028] (2) Then input the LSTM model. LSTM uses its gating mechanism (input gate) Forgotten Gate Output gate This unit can capture temporal dependencies and contextual information in text, learn deep semantic representations of the text, and output a fixed-dimensional semantic feature vector. Its core formula is:
[0029] in, It is the sigmoid function, * is element-wise multiplication, and W and b are model parameters.
[0030] Take the output of the last time step Alternatively, the outputs at all time steps can be averaged / max-pooled to form the semantic feature vector of the entire text. .
[0031] The non-text structured feature extraction submodule is used to extract non-text structured features, including time features, numerical features, and category features. Time features include publication time and deadline, and can be further calculated to include derived time features such as project cycle and remaining time. Numerical features include budget amount, which is normalized. Category features include publishing organization, geographical location, and project type (e.g., engineering, service, procurement), which are then one-hot encoded. In this embodiment, these structured features, together with the semantic feature vector extracted by LSTM, constitute a complete feature representation of the bidding information and are stored in a feature vector database. Figure 1 In the H section, the specific structured features that need to be processed are shown in the following examples: (1) Time characteristics: Calculate the difference between the "release time" and the "deadline time" to obtain the "project cycle"; calculate the difference between the current time and the "deadline time" to obtain the "remaining time". These time characteristics are numerically processed.
[0032] (2) Numerical features: The numerical features such as "budget amount" are normalized by Z-score to make them fall within the range of [0, 1]. The specific formula is as follows:
[0033] (3) Category features: One-hot encoding is used for features with relatively few categories such as "geographic location" and "project type". For category features with a large number of categories such as "publishing organization", an embedding layer is used to map the high-dimensional sparse category features to a low-dimensional dense vector space.
[0034]
[0035] Feature fusion submodule: This module concatenates the text semantic vector output by the Long Short-Term Memory (LSTM) network model with non-textual structured features to form a comprehensive project key feature vector. The text semantic vector... and vectors after processing all non-textual structured features Concatenate the data to form a comprehensive feature vector for the project. And stored in the feature vector database ( Figure 1 (H in the text).
[0036] The intelligent matching and recommendation module is used for matching and recommending based on user profiles and key feature vectors of projects. This matching and recommendation utilizes the deep learning-based recommendation model Wide&Deep, taking key feature vectors of projects and user profiles as input to generate a personalized list of bidding information recommendations. This solution uses the deep learning-based recommendation model Wide&Deep, incorporating semantic features extracted by LSTM (RNN) combined with non-textual structured features into the Wide&Deep model to generate a personalized list of bidding information recommendations, improving the accuracy and diversity of recommendations.
[0037] The recommendation model Wide&Deep includes a Wide part and a Deep part; The Wide part (memory capacity) uses cross-product feature transformations to remember explicit relationships between sparse features. In this scheme, new features are formed through cross-product feature transformations of user features and bidding information features; these are then directly input into the model. This part can effectively capture the combination relationships between features, used to remember historical, explicit feature interactions. For example, "User's industry = Internet and IT" and "Project type = Software development" are cross-product transformed to form a combined feature. If this combination appears frequently in historical click data, the Wide part will assign it a high weight. The formula is: ,in x It is the original feature, / phi(x) This is a cross-transformation feature.
[0038] The Deep part (generalization ability) comprises a deep neural network (DNN) consisting of multiple fully connected layers, used to learn implicit feature representations and complex nonlinear relationships between features. Its input layer receives and concatenates text semantic vectors from the text semantic understanding module, as well as non-text structured features processed by the non-text structured feature extraction module. The recommendation model Wide&Deep merges the outputs of the Wide and Deep parts, and outputs the user's preference probability for bidding information (e.g., whether the user will click on the project) through the Sigmoid activation function; and sorts the preference probabilities of bidding information to obtain a recommendation list of bidding information.
[0039] The Wide & Deep recommendation model is trained by using historical user behavior data (such as user clicks and bidding records) as labels. The parameters of the Wide and Deep parts are jointly optimized through the backpropagation algorithm to minimize the loss function "cross-entropy loss".
[0040] The Wide&Deep recommendation model is implemented as follows: All features, including text semantic vectors, are input into a multilayer feedforward neural network (DNN).
[0041]
[0042]
[0043] After multiple nonlinear transformations (ReLU activation function), a high-order deep feature representation is obtained. .
[0044] Output layer: The output of the Wide layer and the final output of the Deep layer are concatenated, and the probability of user u clicking / bidding on item i is calculated using a Sigmoid function.
[0045]
[0046] in .
[0047] Training: Historical user behavior data (such as user characteristics, bidding information characteristics) is used as labels to construct samples (user, item, label). label=1 indicates that the user has engaged in positive feedback behavior such as clicking / bidding; label=0 indicates exposure without clicking or randomly sampled negative samples. Optimizers such as Adam and SGD are used to jointly train all parameters of the Wide and Deep parts through backpropagation, minimizing the loss function "cross-entropy loss," whose formula is as follows:
[0048] Recommendation generation: For the target user, the model calculates the probability of matching with all candidate items. Sort the results by probability score from highest to lowest to generate a Top-N personalized recommendation list.
[0049] Information push module ( Figure 1 J): Used to push the recommended list of bidding information to users according to a preset push strategy; the push strategy can be set according to user preferences, such as push frequency and push method. The push frequency includes real-time push, daily / weekly summary push, priority push of projects with near deadlines, etc. Highly matched bidding information is then pushed through the user-specified channel (…). Figure 1 The "K" in this context refers to information such as system notifications, emails sent from the registered email address, SMS messages, or app push notifications, which are promptly pushed to users or displayed on the user interface (such as the recommendation page of an app or website).
[0050] Feedback and Model Optimization Module: This module collects user feedback data on the bidding information recommendation list (such as clicks, favorites, bidding intentions, and bidding results). This feedback data is used to iteratively update and optimize the Long Short-Term Memory (LSTM) network model and the Wide&Deep recommendation model, enabling the system to learn and continuously improve. The system continuously collects user feedback on the pushed information. For example, a user might click to view, favorite, mark as interested, or report as irrelevant for a project. If the user ultimately participates in the project and wins the bid, this feedback carries higher weight. (Feedback and Model Optimization Module...) Figure 1L) collects these feedback data (S108) as new training samples. This data will be used for: Optimize the LSTM model: Adjust the weights of the LSTM model through the backpropagation algorithm to make it better able to capture the deep semantic features of bidding information that users are interested in.
[0051] Optimize the recommendation model: Update the recommendation algorithm model to make the generated recommendation list more consistent with the user's actual preferences.
[0052] Update user profiles: Dynamically adjust and improve user profiles based on user behavior and feedback to make them more accurate.
[0053] Through this closed-loop feedback and optimization mechanism, the entire system can achieve continuous learning and performance improvement.
[0054] A method for intelligent bidding information push based on RNN and Wide&Deep, such as Figure 2 As shown, it includes the following steps: S101. Data Acquisition and Preprocessing Steps: Collect bidding information and preprocess the collected bidding information; the data acquisition and preprocessing module automatically collects bidding information through various channels, including major bidding platforms, government announcements, and industry websites. The collected bidding information includes, but is not limited to, project name, issuing agency, release time, deadline, project summary, technical requirements, budget amount, and geographical location; the preprocessing includes cleaning, deduplication, word segmentation, part-of-speech tagging, and removal of stop words from the text data.
[0055] S102. User profile construction steps: Obtain user data and construct multi-dimensional user profiles based on the user data; in the user profile construction steps, the user data includes the user's registration information, browsing history, search records, click behavior, winning project records, enterprise qualifications, industry preferences, and geographical location preferences; the user profile includes the user's industry, technology field, project type, budget range, geographical region, and institutions they have cooperated with.
[0056] The bidding information feature extraction steps are as follows: Feature vectors are extracted from the preprocessed bidding text using a Long Short-Term Memory (LSTM) network model based on an RNN to obtain a text semantic vector; this is also used to extract non-textual structured features from the bidding text; furthermore, the text semantic vector output by the LSTM network model is concatenated with the non-textual structured features to form a key project feature vector; the bidding information feature extraction steps include the following sub-steps: S103, Deep Semantic Understanding Sub-step: Using the Long Short-Term Memory (LSTM) network model, sequence modeling is performed on the preprocessed bidding text; firstly, the Word2Vec word vector model is used to convert the words in the bidding text into low-dimensional dense vector representations, which are then used as input to the LSTM network model. S104, Non-textual structured feature extraction sub-step: Extract non-textual structured features, including time features, numerical features, and category features; Time features: release time, deadline, and can further calculate derived time features such as project cycle and remaining time; Numerical features: budget amount, and perform normalization processing; Category features: publishing organization, geographical location, project type; Feature fusion sub-step: Concatenate the text semantic vector with the non-text structured features to form a comprehensive project key feature vector.
[0057] S105, Intelligent Matching and Recommendation Steps: Matching and recommendation are performed based on user profiles and key feature vectors of projects; the matching and recommendation uses the deep learning-based recommendation model Wide&Deep, taking the key feature vectors of projects and user profiles as inputs to the Wide&Deep recommendation model to generate a personalized list of bidding information recommendations; this solution uses a deep learning-based recommendation model (Wide&Deep), which integrates the semantic features extracted by RNN with non-textual structured features into the recommendation model to generate a personalized list of bidding information recommendations, improving the accuracy and diversity of recommendations.
[0058] The recommendation model Wide&Deep includes a Wide part and a Deep part; The Wide part (memory capacity) employs cross-product feature transformations to remember explicit relationships between sparse features. In this scheme, new features are formed through cross-product feature transformations of user features and bidding information features; these are then directly input into the model. This part effectively captures the combinatorial relationships between features, enabling the memorization of historical, explicit feature interactions.
[0059] The Deep part (generalization ability) comprises a deep neural network (DNN) consisting of multiple fully connected layers, used to learn implicit feature representations and complex nonlinear relationships between features. Its input layer receives and concatenates text semantic vectors from the text semantic understanding module, as well as non-text structured features processed by the non-text structured feature extraction module. The recommendation model Wide&Deep fuses the outputs of the Wide and Deep parts, and outputs the user's preference probability for bidding information (e.g., whether the user will click on the project) through a Sigmoid activation function; then, it sorts the preference probabilities of the bidding information to obtain a recommendation list of bidding information. The recommendation list of bidding information is as follows: S106. Information Push Steps: The recommended list of bidding information is pushed to users according to a preset push strategy; the push strategy can be set according to user preferences, such as push frequency and push method. The push channels / methods of the information push module include in-system messages, emails, SMS, and App pushes.
[0060] Feedback and Model Optimization Steps: Collect user feedback data on the bidding information recommendation list and use this data to iteratively update and optimize the Long Short-Term Memory (LSTM) network model and the Wide&Deep recommendation model. The system continuously collects user feedback on the pushed information (S107). For example, a user may click to view, favorite, mark as interested, or report as irrelevant for a project. If the user ultimately participates in the project and wins the bid, the feedback carries a higher weight. The feedback and model optimization steps collect this feedback data (S108) as new training samples. This data will be used for: Optimize the LSTM model (S109): Adjust the weights of the LSTM model through the backpropagation algorithm to make it better able to capture the deep semantic features of bidding information that users are interested in.
[0061] Optimize the recommendation model (S109): Update the recommendation algorithm model so that the generated recommendation list is more in line with the user's real preferences.
[0062] Update user profile (S109): Dynamically adjust and improve user profiles based on user behavior and feedback to make them more accurate.
[0063] Through this closed-loop feedback and optimization mechanism, the entire system can achieve continuous learning and performance improvement.
[0064] The above descriptions are merely embodiments of the present invention. Commonly known structures and characteristics are not described in detail here. Those skilled in the art are aware of all common technical knowledge in the field prior to the application date or priority date, are aware of all existing technologies in that field, and have the ability to apply conventional experimental methods prior to that date. Those skilled in the art can, based on the guidance provided in this application, improve and implement this solution in combination with their own capabilities. Some typical known structures or methods should not be obstacles for those skilled in the art to implement this application. It should be noted that those skilled in the art can make several modifications and improvements without departing from the structure of the present invention. These should also be considered within the scope of protection of the present invention, and will not affect the effectiveness of the implementation of the present invention or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.
Claims
1. An intelligent bidding information push system based on RNN and Wide & Deep, characterized in that, include: Data acquisition and preprocessing module: used to collect bidding information and preprocess the collected bidding information; User profile building module: used to acquire user data and build multi-dimensional user profiles based on the user data; The bidding information feature extraction module is used to extract feature vectors from preprocessed bidding texts using the Long Short-Term Memory (LSTM) network model based on RNN to obtain text semantic vectors; it is also used to extract non-textual structured features from bidding texts; and it is also used to concatenate the text semantic vectors output by the LSTM network model with the non-textual structured features to form the key feature vectors of the project. Intelligent matching and recommendation module: used for matching and recommending based on user profiles and key feature vectors of projects; the matching and recommendation uses the deep learning-based recommendation model Wide & Deep, taking the key feature vectors of projects and user profiles as inputs to the recommendation model Wide & Deep, to generate a personalized list of bidding information recommendations; Information push module: used to push a list of recommended bidding information to users according to a preset push strategy; Feedback and Model Optimization Module: This module collects user feedback data on the recommended list of bidding information and uses the feedback data to iteratively update and optimize the Long Short-Term Memory (LSTM) network model and the recommendation model Wide & Deep.
2. The intelligent bidding information push system based on RNN and Wide & Deep as described in claim 1, characterized in that: The data acquisition and preprocessing module automatically collects bidding information through various channels, including major bidding platforms, government announcements, and industry websites. The collected bidding information includes, but is not limited to, project name, issuing agency, release time, deadline, project summary, technical requirements, budget amount, and geographical location. The preprocessing includes cleaning, deduplication, word segmentation, part-of-speech tagging, and removal of stop words from the text data.
3. The intelligent bidding information push system based on RNN and Wide & Deep as described in claim 2, characterized in that: In the user profile building module, the user data includes the user's registration information, browsing history, search records, click behavior, winning bid records, enterprise qualifications, industry preferences, and geographical location preferences; The user profile includes the industry, technology field, project type, budget range, geographical region, and institutions the user has cooperated with.
4. The intelligent bidding information push system based on RNN and Wide & Deep as described in claim 3, characterized in that: The bidding information feature extraction module Includes the following sub-modules: The deep semantic understanding submodule utilizes the Long Short-Term Memory (LSTM) network model to perform sequence modeling on the preprocessed bidding text. Specifically, the Word2Vec word vector model is first used to convert the words in the bidding text into low-dimensional dense vector representations, which serve as the input to the LSTM network model. Non-text structured feature extraction submodule: used to extract non-text structured features, including time features, numerical features, and category features; time features: release time, deadline, and can further calculate derived time features such as project cycle and remaining time; numerical features: budget amount, and normalization processing; category features: publishing organization, geographical location, and project type; Feature fusion submodule: concatenates text semantic vectors with non-text structured features.
5. The intelligent bidding information push system based on RNN and Wide & Deep as described in claim 4, characterized in that: The recommendation model Wide & Deep includes a Wide part and a Deep part; The Wide part is used to form new features through cross-feature transformation of user features and bidding information features; The Deep part contains a deep neural network (DNN) consisting of multiple fully connected layers. Its input layer is used to receive and concatenate text semantic vectors from the text semantic understanding module and non-text structured features processed by the non-text structured feature extraction module. The recommendation model Wide & Deep fuses the outputs of the Wide and Deep parts and outputs the user's preference probability for bidding information through the Sigmoid activation function. The preference probability of bidding information is sorted to obtain a recommended list of bidding information.
6. The intelligent bidding information push system based on RNN and Wide & Deep as described in claim 5, characterized in that: The information push module's push channels include in-system messages, emails, SMS messages, and App push notifications.
7. A method for intelligent bidding information push based on RNN and Wide & Deep, characterized in that: Includes the following steps: Data collection and preprocessing steps: Collect bidding information and preprocess the collected bidding information; User profile building steps: Obtain user data and build multi-dimensional user profiles based on the user data; The steps for extracting features from bidding information are as follows: The preprocessed bidding text is extracted using the Long Short-Term Memory (LSTM) network model based on RNN to obtain a text semantic vector; it is also used to extract non-textual structured features from the bidding text; and it is also used to concatenate the text semantic vector output by the LSTM network model with the non-textual structured features to form the key feature vector of the project. Intelligent matching and recommendation steps: Matching and recommendation are performed based on user profiles and key feature vectors of projects; the matching and recommendation uses the deep learning-based recommendation model Wide & Deep, which takes key feature vectors of projects and user profiles as inputs to the Wide & Deep recommendation model to generate a personalized list of bidding information recommendations; Information push steps: Push the list of recommended bidding information to users according to the preset push strategy; Feedback and model optimization steps: Collect user feedback data on the bidding information recommendation list, and use the feedback data to iteratively update and optimize the Long Short-Term Memory (LSTM) network model and the recommendation model Wide & Deep.
8. The intelligent bidding information push method based on RNN and Wide & Deep as described in claim 7, characterized in that: The bidding information feature extraction step includes the following sub-steps: The deep semantic understanding sub-step of the text is as follows: The preprocessed bidding text is sequence modeled using the Long Short-Term Memory (LSTM) network model. First, the Word2Vec word vector model is used to convert the words in the bidding text into low-dimensional dense vector representations, which are then used as input to the LSTM network model. Non-text structured feature extraction sub-step: Extract non-text structured features, which include time features, numerical features, and category features; Time features: release time, deadline, and can be further calculated to obtain derived time features such as project cycle and remaining time; Numerical features: budget amount, which is normalized; Category features: publishing organization, geographical location, and project type; Feature fusion sub-step: concatenate the text semantic vector with the non-text structured features.
9. The intelligent bidding information push method based on RNN and Wide & Deep as described in claim 8, characterized in that: The recommendation model Wide & Deep includes a Wide part and a Deep part; The Wide part is used to form new features through cross-feature transformation of user features and bidding information features; The Deep part contains a deep neural network (DNN) consisting of multiple fully connected layers. Its input layer is used to receive and concatenate text semantic vectors from the text semantic understanding module and non-text structured features processed by the non-text structured feature extraction module. The recommendation model Wide & Deep fuses the outputs of the Wide and Deep parts and outputs the user's preference probability for bidding information through the Sigmoid activation function. The preference probability of bidding information is sorted to obtain a recommended list of bidding information.