Driver recommendation system based on multimodal feature fusion

Abstract

The present invention relates to the field of information technology, and specifically to a driver recommendation system based on multimodal feature fusion, including a basic data layer for collecting basic data about drivers and cargo sources; a data processing layer connected to the basic data layer, including a data processing module for preprocessing and feature extraction of basic data to obtain multimodal feature vectors; a feature fusion module for weighted fusion of feature vectors; a first recall module for recalling feature vectors; a sorting model for sorting coarse recall features to obtain sorting feature data; and a recommendation business layer, the recommendation business layer including a second recall module for recalling feature vectors to obtain candidate set data. The present invention comprehensively utilizes deep learning, attention mechanism and multimodal feature fusion to conduct in-depth analysis of drivers' historical behaviors, real-time status and cargo source characteristics, and provides drivers with personalized and efficient cargo source recommendations.

Description

Technical Field

[0001] The present invention relates to the field of information technology, and in particular to a driver recommendation system based on multimodal feature fusion. Background Art

[0002] In the modern logistics industry, vehicle-cargo matching is a core issue. Traditional vehicle-cargo matching methods usually rely on matching methods based on rules and labels, which have many problems in practical applications: 1. Lack of flexibility. Existing label matching methods can only consider preset specific features, such as cargo type and transportation distance, and cannot dynamically adapt to changes in driver needs. 2. Lack of diversity in recommendation results: Label matching usually produces similar recommendation results, which may cause drivers to miss potential transportation opportunities. 3. Difficulty in responding to market changes: As market demand and driver preferences continue to change, traditional matching methods are difficult to adapt and adjust quickly. Summary of the invention

[0003] The purpose of the present invention is to provide a driver recommendation system based on multimodal feature fusion to solve the above technical problems;

[0004] The technical problem solved by the present invention can be achieved by adopting the following technical solutions:

[0005] Driver recommendation system based on multimodal feature fusion, including:

[0006] The basic data layer is used to collect basic data about drivers and cargo sources;

[0007] The data processing layer, connected to the basic data layer, includes:

[0008] A data processing module, connected to the basic data layer, for preprocessing and extracting features from the basic data to obtain a multi-modal feature vector;

[0009] A feature fusion module, connected to the data processing module, for performing weighted fusion on the feature vectors to obtain a feature fusion result that combines the driver portrait and the cargo source portrait;

[0010] A first recall module, connected to the data processing module, for recalling the feature vector to obtain a rough recall feature;

[0011] A sorting model, connected to the feature fusion module, for sorting the coarse recall features according to the feature fusion result to obtain sorting feature data;

[0012] A recommendation business layer is connected to the data processing layer, and the recommendation business layer includes a second recall module, which is used to recall the feature vector to obtain candidate set data. The recommendation business layer outputs a recommendation result including the sorting feature data and the candidate set data.

[0013] Preferably, the basic data includes at least the driver's historical order grabbing records, cargo source characteristic data, real-time status data, driver sequence data and driver geographic media data.

[0014] Preferably, the data processing module preprocesses the structured data of the basic data by maximum and minimum normalization, and the normalization expression is:

[0015]

[0016] Among them, x * represents the normalized result, x represents the original structured data, and x min Represents the minimum value of the structured data, x max Indicates the maximum value of the structured data.

[0017] Preferably, the data processing module extracts features from the unstructured text data of the basic data through a pre-trained language model, and the expression is:

[0018] E text =BERT(T)

[0019] Where T represents the input text, BERT represents the language model, and E text Represents the extracted text feature vector;

[0020] The data processing module extracts the temporal characteristics of the sequence data in the basic data through the long short-term memory network, and the expression is:

[0021] h t =LSTM(x t ,h t-1 )

[0022] Among them, x t represents the current input in the sequence data, h t represents the current hidden state, h t-1 Indicates the hidden state at the previous moment.

[0023] Preferably, the feature fusion module performs weighted fusion on the feature vectors through an attention mechanism, and the expression is:

[0024]

[0025] Among them, h irepresents the feature representation of the i-th mode, n represents the number of modal features, α i Represents the weight calculated by the attention mechanism.

[0026] Preferably, the first recall module calculates the similarity matrix between drivers through collaborative filtering and obtains the rough recall feature. The calculation formula of the similarity matrix is:

[0027]

[0028] Among them, r ui and r vi Represent the ratings of user u and user v on item i, respectively. represents the average share of all items by user u, represents the average score of user v on all items, I uv Represents the set of items rated by users u and v.

[0029] Preferably, the recommendation business layer includes a cache database for caching the sorting feature data and the candidate set data, and the recommendation business layer also adjusts the recommendation results according to the driver portrait, the source portrait and the similarity matrix.

[0030] Preferably, the sorting model uses the Wide&Deep model and the GBDT model to sort the coarse recall features, and the loss function of the sorting model is:

[0031]

[0032] Among them, y i represents the true label, p(y i |x i ; θ) represents the model prediction probability, θ represents the model parameters, and N represents the number of labels.

[0033] Preferably, the second recall module includes:

[0034] A collaborative filtering recall unit, configured to output first recommended data according to the similarity matrix;

[0035] A content similarity recall unit, used to calculate the content similarity between the source orders and output the second recommended data;

[0036] A sequence recall unit, used to analyze the driver's transportation behavior sequence through a sequence model, predict the driver's transportation demand, and output third recommendation data;

[0037] The candidate set data is obtained by merging the first recommended data, the second recommended data and the third recommended data.

[0038] Preferably, it also includes,

[0039] A model optimization module, connected to the ranking model, for training and optimizing the recommendation model using historical data;

[0040] The recommendation feedback module is connected to the recommendation business layer and is used to obtain feedback data from drivers and update the driver portrait and the recommendation model based on the feedback data.

[0041] Beneficial effects of the present invention: Due to the adoption of the above technical scheme, the present invention comprehensively utilizes deep learning, attention mechanism and multimodal feature fusion to conduct in-depth analysis of the driver's historical behavior, real-time status and cargo source characteristics, and provide drivers with personalized and efficient cargo source recommendations. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 Schematic diagram of the driver recommendation system based on multimodal feature fusion in an embodiment of the present invention.

[0043] In the attached figure: 1. Basic data layer; 11. Business database; 2. Data processing layer; 21. Data processing module; 22. Feature fusion module; 23. First recall module; 24. Sorting model; 3. Recommendation business layer; 31. Second recall module; 311. Content similarity recall unit; 312. Collaborative filtering recall unit; 313. Sequence recall unit; 32. Cache database; 33. Recommendation center. DETAILED DESCRIPTION

[0044] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0045] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments may be combined with each other.

[0046] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments, but they are not intended to limit the present invention.

[0047] Driver recommendation system based on multimodal feature fusion, such as Figure 1 As shown, including

[0048] Basic data layer 1, used to collect basic data about drivers and cargo sources;

[0049] Data processing layer 2, connected to basic data layer 1, includes:

[0050] The data processing module 21 is connected to the basic data layer 1 and is used to preprocess and extract features of the basic data to obtain multi-modal feature vectors;

[0051] The feature fusion module 22 is connected to the data processing module 21 and is used to perform weighted fusion on the feature vectors to obtain a feature fusion result that combines the driver portrait and the cargo source portrait;

[0052] A first recall module 23, connected to the data processing module 21, is used to recall the feature vector to obtain a rough recall feature;

[0053] The sorting model 24 is connected to the feature fusion module 22 and is used to sort the coarse recall features according to the feature fusion results to obtain sorting feature data;

[0054] The recommendation business layer 3 is connected to the data processing layer 2. The recommendation business layer 3 includes a second recall module 31. The second recall module 31 is used to recall the feature vector to obtain the candidate set data. The recommendation business layer 3 outputs the recommendation result including the sorted feature data and the candidate set data.

[0055] Specifically, the present invention proposes an intelligent driver recommendation system based on multimodal feature fusion, which analyzes the multi-dimensional characteristics of drivers and combines multiple deep learning models to achieve more flexible and efficient cargo source recommendations.

[0056] First, the basic data of drivers and cargo sources are collected and stored through the basic data layer 1, including the driver's historical order grabbing records, cargo source characteristic data, driver characteristic data, etc.

[0057] The data processing layer 2 preprocesses the collected data, including steps such as data cleaning, feature extraction and data standardization.

[0058] The feature fusion module 22 is based on deep learning models such as Transformer and LSTM, and integrates multiple data sources to perform feature learning and prediction.

[0059] The recommendation business layer 3 generates candidate set data based on the processed data using a variety of recommendation algorithms (such as collaborative filtering, content similarity, sequence modeling, etc.).

[0060] The present invention also includes an online recommendation layer for real-time monitoring of the driver's operating behavior and location data, combining the recommendation list generated by the recommendation service layer 3, and dynamically adjusting the recommendation results.

[0061] It also includes a model training and optimization layer, which regularly trains and optimizes the recommendation model to improve the accuracy and response speed of recommendations.

[0062] The present invention comprehensively utilizes deep learning, attention mechanism and multimodal feature fusion to conduct in-depth analysis of the driver's historical behavior, real-time status and cargo source characteristics. The generated recommendation results are sent to the driver's application interface through the recommendation center 23, providing the driver with personalized and efficient cargo source recommendations.

[0063] Specifically, the data sources of the second recall module 31 and the sorting model 24 in the present invention are both from the feature vectors obtained by the data processing layer 2 . Usually, the amount of data obtained by the sorting model 24 is greater than the amount of data obtained by the second recall module 31 .

[0064] In a preferred embodiment, the basic data includes at least the driver's historical order grabbing records, cargo source characteristic data, real-time status data, driver sequence data and driver geographic media data.

[0065] Specifically, the basic data layer 1 collects and stores basic data of drivers and cargo sources, including drivers' historical order grabbing records, cargo source characteristic data, driver characteristic data, etc.

[0066] In this embodiment, the basic data layer 1 includes a business database 11, specifically a doris database. When processing large-scale data, Doris can support fast real-time queries and multiple concurrent requests. Doris is used as a database unit for storing and querying a large amount of driver, cargo source and other data, which can ensure that the data processing layer 2 can access and process data quickly and stably.

[0067] More specifically, the basic data layer 1 obtains the driver's historical behavior data and cargo source characteristic data through the logistics platform, including:

[0068] The driver's historical record of grabbing orders, such as the number of orders grabbed, the time of grabbing orders, etc.

[0069] Cargo source characteristic data, such as cargo type, transportation distance, loading and unloading location, etc.;

[0070] Real-time status data, such as the driver’s current location, equipment information, etc.

[0071] Driver sequence data, such as order grabbing sequence data;

[0072] Driver geo-media data, such as driver time, geographic location, social media interaction information, etc.

[0073] In a preferred embodiment, the data processing module 21 preprocesses the structured data of the basic data by maximum and minimum normalization, and the normalization expression is:

[0074]

[0075] Among them, x * represents the normalized result, x represents the original structured data, and xmin Represents the minimum value of structured data, x max Represents the maximum value of structured data.

[0076] In a preferred embodiment, the data processing module 21 extracts features from the unstructured text data of the basic data through a pre-trained language model, and the expression is:

[0077] E text =BERT(T)

[0078] Where T represents the input text, BERT represents the language model, and E text Represents the extracted text feature vector;

[0079] The data processing module 21 extracts the temporal characteristics of the sequence data in the basic data through the long short-term memory network, and the expression is:

[0080] h t =LSTM(x t ,h t-1 )

[0081] Among them, x t Represents the current input in the sequence data, h t represents the current hidden state, h t-1 Represents the hidden state at the previous moment.

[0082] The multimodal data of drivers collected by the present invention includes:

[0083] Structured data, such as basic attributes of drivers;

[0084] Unstructured text data, such as driver reviews, order feedback, etc.

[0085] Unstructured user behavior sequence data, such as historical transportation routes, order completion time, etc.

[0086] It also includes real-time data such as the driver's current location, device information, network status, or image data.

[0087] The collected data is stored in a business database 11 (such as Doris) to ensure data persistence and availability. Data redundancy and backup mechanisms are used to ensure data security and integrity.

[0088] Specifically, the data processing layer 2 first cleans the basic data to remove abnormal values ​​and duplicate values ​​in the data to ensure the accuracy of the data.

[0089] The data is normalized to eliminate the dimensional differences between different features. Use the Min-Max normalization method:

[0090]

[0091] Among them, x * represents the normalized result, x represents the original structured data, and x min Represents the minimum value of structured data, x max Represents the maximum value of structured data.

[0092] Unstructured text data processing:

[0093] Use pre-trained models such as BERT to extract features from unstructured text data and generate high-dimensional vector representations:

[0094] E text =BERT(T)

[0095] Among them, T represents the input text, E text is the extracted text feature vector

[0096] Unstructured sequence data processing:

[0097] Use LSTM (Long Short-Term Memory Network) to process the driver's historical behavior sequence data and extract behavior sequence features:

[0098] h t =LSTM(x t ,h t-1 )

[0099] Among them, x t represents the current input in the sequence, h t represents the current hidden state, h t-1 Indicates the hidden state at the previous moment.

[0100] In a preferred embodiment, the feature fusion module 22 performs weighted fusion on the feature vectors through an attention mechanism, and the expression is:

[0101]

[0102] Among them, h i represents the feature representation of the i-th mode, n represents the number of modal features, α i Represents the weight calculated by the attention mechanism.

[0103] Specifically, the multimodal feature fusion of the present invention uses the attention mechanism to perform weighted fusion of features of different modalities, and the formula is as follows:

[0104]

[0105] Among them, hi is the characteristic representation of the i-th mode, α i is the weight calculated by the attention mechanism

[0106] Specifically, multi-dimensional features are extracted from driver and cargo data, such as time, geographic location, social media interaction information, etc.

[0107] The attention mechanism in Transformer is used to perform weighted fusion of multimodal features (including text, images, and behavioral data).

[0108] Specifically, for features of different modalities, for example, features of text data are extracted through the BERT model, features of image data are extracted through convolutional neural networks (CNN), and time-dependent features of behavioral data are extracted through sequence modeling.

[0109] Then, Transformer uses the attention mechanism to weight the features of these different modalities, assigning different weights to highlight key features, and finally fusing them into a unified high-dimensional feature representation, improving the accuracy of the feature representation and the predictive ability of the model.

[0110] In a preferred embodiment, the first recall module 23 calculates the similarity matrix between drivers through collaborative filtering and obtains the rough recall feature. The calculation formula of the similarity matrix is:

[0111]

[0112] Among them, r ui and r vi Represent the ratings of user u and user v on item i, respectively. represents the average share of all items by user u, represents the average score of user v on all items, I uv Represents the set of items rated by user u and user v.

[0113] This formula calculates the Pearson Correlation Coefficient, which is used to measure the similarity of rating behaviors between two users (or drivers).

[0114] The molecular part

[0115]

[0116] The sum of the products of the deviations of user u and user v on the items they rated together is calculated.

[0117] Each product term represents the degree to which the ratings of user u and user v deviate from their respective average ratings for an item i.

[0118] This product term is positive if the two ratings diverge in the same direction (for example, both users give an item a high rating); it is negative if they diverge in opposite directions (one user gives a high rating and the other user gives a low rating).

[0119] Denominator

[0120]

[0121] It calculates the square root of the sum of the squares of the rating deviations of user u and user v on the commonly rated items.

[0122] For multiple users, the present invention adopts a combination method to select two users in turn as user u and user v for comparison, and then calculates and outputs matrixed results for multiple times to obtain a similarity matrix.

[0123] In a preferred embodiment, the recommendation business layer 3 includes a cache database 32 for caching sorting feature data and candidate set data. The recommendation business layer 3 also adjusts the recommendation results according to the driver portrait, the source portrait and the similarity matrix.

[0124] The sorted recommendation results are cached in a cache database 32, such as Redis, and dynamically adjusted according to the real-time behavior and status of the driver to ensure the real-time and accuracy of the recommendation results.

[0125] Specifically, the present invention dynamically adjusts the recommendation results through real-time data to ensure that the system can respond to market changes and driver needs in a timely manner, and uses SHAP values ​​to explain the recommendation results to improve users' understanding and trust in the system.

[0126] In a preferred embodiment, the sorting model 24 uses the Wide&Deep model and the GBDT model to sort the coarse recall features, and the loss function of the sorting model 24 is:

[0127]

[0128] Among them, y i represents the true label, p(y i |x i ; θ) represents the model prediction probability, θ represents the model parameters, and N represents the number of labels.

[0129] Specifically, the present invention uses the Wide&Deep model and the GBDT model to sort the recalled candidate sets.

[0130] The Wide & Deep model combines a linear model (Wide part) and a deep neural network (Deep part) to capture high-frequency co-occurrence features and nonlinear feature combinations.

[0131] Its loss function is:

[0132]

[0133] Among them, y i is the true label, p(y i |x i ; θ) is the model prediction probability, and θ is the model parameter.

[0134] The purpose of the loss function is to minimize the difference between the model's prediction and the actual label to obtain the optimal parameters. The Wide & Deep model can jointly learn the relationship between users and items in a linear and nonlinear way to perform sorting.

[0135] The GBDT model can also effectively rank candidate items by learning the complex relationship between features and targets.

[0136] In addition to the recommendation results cached in the cache database 32, the present invention also needs to dynamically adjust the recommendation results based on the driver portrait, cargo source portrait and similarity matrix, wherein the driver portrait and cargo source portrait refer to data used to represent driver characteristics and cargo source characteristics. For example, the recommendation results are adjusted according to the driver's current behavior and status, such as the driver's needs, location, historical behavior, etc., to meet the driver's needs.

[0137] Specifically, the present invention dynamically adjusts the recommendation results by real-time monitoring of the driver's operating behavior and location data, combined with the candidate set data generated by the recommendation business layer 3.

[0138] The present invention also includes a feedback function, which updates the driver portrait and recommendation model through the driver's click, browse, grab order and other behavioral feedback, and continuously optimizes the recommendation effect.

[0139] In a preferred embodiment, the second recall module 31 includes:

[0140] A collaborative filtering recall unit 312, configured to output first recommended data according to a similarity matrix;

[0141] A content similarity recall unit 311, configured to calculate the content similarity between the source orders and output second recommended data;

[0142] A sequence recall unit 313, used to analyze the driver's transportation behavior sequence through a sequence model, predict the driver's transportation demand, and output third recommendation data;

[0143] The first recommended data, the second recommended data and the third recommended data are combined to obtain the candidate set data.

[0144] Specifically, the collaborative filtering recall unit 312 uses a collaborative filtering algorithm (UserCF and ItemCF) to calculate the similarity between drivers (UserCF) based on driver behavior data, and recommends sources of goods that similar drivers are interested in.

[0145] Based on the similarity of the cargo list (ItemCF), cargo lists that are similar to the driver’s historical preferences are recommended.

[0146] The content similarity recall unit 311 uses content similarity calculation to calculate the content similarity between the source orders, including the similarity of the loading place, unloading place, cargo type, etc. The source orders with similar content are recommended to the drivers who meet the driver's historical preferences.

[0147] The sequence recall unit 313 adopts a sequence modeling algorithm, and uses a sequence modeling algorithm (such as LSTM, GRU, etc.) to analyze the driver's transportation behavior sequence and predict the driver's future transportation needs.

[0148] The present invention combines collaborative filtering, content similarity and sequence modeling to provide diversified recommendation results and reduce the homogeneity problem of recommendations.

[0149] Specifically, the main task of the recommendation service layer 3 is to generate the final recommendation results. This layer combines the feature output of the data processing layer 2 and uses a variety of recommendation algorithms to generate a recommendation list.

[0150] The recall strategy specifically includes:

[0151] Content similarity recall, calculates the content similarity between source orders, including features such as loading location, unloading location, and cargo type.

[0152] The formula is as follows:

[0153]

[0154] Among them, x ik and x jk Respectively represent the values ​​of item i and j on feature k, w k is the weight of feature k, δ(x ik ,x jk ) is a binary indicator function.

[0155] Collaborative filtering recall: Based on the similarity matrix of drivers, recommend cargo orders that similar drivers are interested in.

[0156] The collaborative filtering process of the present invention uses a user-based collaborative filtering (UserCF) method to calculate the similarity matrix between drivers and generate a preliminary recommendation candidate set.

[0157] The similarity calculation formula is the same as that of the first recall module 23, as follows:

[0158]

[0159] Among them, r ui and r vi Represents the ratings of user u and v on item i, I uv Represents the set of items rated by users u and v.

[0160] Specifically, the goal of the sequence recall unit 313 is to predict future transportation demand based on the driver's historical transportation behavior.

[0161] Use sequence modeling algorithms, such as LSTM (Long Short-Term Memory Network) or GRU (Gated Recurrent Unit), to predict the driver's future needs by analyzing the driver's past transportation behavior data. For example, based on the driver's historical transportation time, frequency, route and other data, predict the source of goods that the driver will choose next, and recommend the source of goods that the driver is interested in in the future.

[0162] In the second recall module 31, the three recall results (collaborative filtering recommendation, content similarity recommendation, and sequence recommendation) are merged into a candidate set data, and the candidate set data and the sorting feature data output by the sorting model 24 are used together as the final recommendation result;

[0163] The recommendation results can be recommended to the driver in the form of a list through the recommendation center 33. The present invention effectively avoids the uniformity of the recommendation results through a variety of recall strategies, thereby improving the accuracy of the recommendation and user experience, and achieving the recommendation of the most suitable sources of goods for the driver while improving the diversity and accuracy of the recommendation, reducing the problem of homogeneous recommendations.

[0164] In a preferred embodiment, it also includes:

[0165] A model optimization module, connected to the ranking model 24, is used to train and optimize the recommendation model using historical data;

[0166] The recommendation feedback module is connected to the recommendation business layer 3 and is used to obtain the driver's feedback data and update the driver portrait and recommendation model based on the feedback data.

[0167] Specifically, the present invention supports real-time monitoring and recommendation adjustment, and can dynamically adjust the recommendation results according to the real-time status of the driver (such as current location, time, equipment information, etc.) to ensure the real-time and accuracy of the recommendation.

[0168] The model optimization module uses historical data to train the recommendation model, optimize model parameters, and improve the model's predictive ability. It also regularly evaluates the model's recommendation effect, adjusts the model structure and parameters, and improves the accuracy and diversity of recommendations.

[0169] Specifically, the present invention uses SHAP (Shapley Additive Explanations) value to explain the recommendation results, improve the transparency of the system and user trust, collect drivers' feedback on the recommendations (such as clicks, browsing, grabbing orders, etc.), and continuously optimize the driver portrait and recommendation model.

[0170] The above description is only a preferred embodiment of the present invention, and does not limit the implementation mode and protection scope of the present invention. For those skilled in the art, it should be aware that all solutions obtained by equivalent substitutions and obvious changes made using the description and illustrations of the present invention should be included in the protection scope of the present invention.

Claims

1. Driver recommendation system based on multimodal feature fusion, characterized by: include, Basic data layer (1), used to collect basic data about drivers and cargo sources; The data processing layer (2), connected to the basic data layer (1), comprises: A data processing module (21), connected to the basic data layer (1), for preprocessing and extracting features from the basic data to obtain a multi-modal feature vector; A feature fusion module (22), connected to the data processing module (21), is used to perform weighted fusion on the feature vectors to obtain a feature fusion result that combines the driver portrait and the cargo source portrait; A first recall module (23), connected to the data processing module (21), for recalling the feature vector to obtain a rough recall feature; A sorting model (24), connected to the feature fusion module (22), for sorting the coarse recall features according to the feature fusion result to obtain sorting feature data; A recommendation business layer (3) is connected to the data processing layer (2), and the recommendation business layer (3) includes a second recall module (31). The second recall module (31) is used to recall the feature vector to obtain candidate set data. The recommendation business layer (3) outputs a recommendation result including the sorted feature data and the candidate set data.

2. The driver recommendation system based on multimodal feature fusion according to claim 1, characterized in that: The basic data at least includes the driver's historical order grabbing records, cargo source characteristic data, real-time status data, driver sequence data and driver geographic media data.

3. The driver recommendation system based on multimodal feature fusion according to claim 1, characterized in that: The data processing module (21) pre-processes the structured data of the basic data by maximum and minimum normalization, and the normalization expression is: Among them, x * represents the normalized result, x represents the original structured data, and x min Represents the minimum value of the structured data, x max Indicates the maximum value of the structured data.

4. The driver recommendation system based on multimodal feature fusion according to claim 1, characterized in that: The data processing module (21) extracts features from the unstructured text data of the basic data through a pre-trained language model, and the expression is: E text =BERT(T) Where T represents the input text, BERT represents the language model, and E text Represents the extracted text feature vector; The data processing module (21) extracts the temporal characteristics of the sequence data in the basic data through a long short-term memory network, and the expression is: h t =LSTM(x t ,h t-1 ) Among them, x t represents the current input in the sequence data, h t represents the current hidden state, h t-1 Represents the hidden state at the previous moment.

5. The driver recommendation system based on multimodal feature fusion according to claim 1, characterized in that: The feature fusion module (22) performs weighted fusion on the feature vectors through an attention mechanism, and the expression is: Among them, h i represents the feature representation of the i-th mode, n represents the number of modal features, α i Represents the weight calculated by the attention mechanism.

6. The driver recommendation system based on multimodal feature fusion according to claim 1, characterized in that: The first recall module (23) calculates the similarity matrix between drivers through collaborative filtering and obtains the rough recall feature. The calculation formula of the similarity matrix is: Among them, r ui and r vi Represent the ratings of user u and user v on item i, respectively. represents the average share of all items by user u, represents the average score of user v on all items, I uv Represents the set of items rated by users u and v.

7. The driver recommendation system based on multimodal feature fusion according to claim 6, characterized in that: The recommendation business layer (3) includes a cache database (32) for caching the sorting feature data and the candidate set data. The recommendation business layer (3) also adjusts the recommendation result according to the driver portrait, the cargo source portrait and the similarity matrix.

8. The driver recommendation system based on multimodal feature fusion according to claim 1, characterized in that: The sorting model (24) uses the Wide&Deep model and the GBDT model to sort the coarse recall features. The loss function of the sorting model (24) is: Among them, y i represents the true label, p(y i |x i ; θ) represents the model prediction probability, θ represents the model parameters, and N represents the number of labels.

9. The driver recommendation system based on multimodal feature fusion according to claim 6, characterized in that: The second recall module (31) comprises: A collaborative filtering recall unit (312), configured to output first recommended data according to the similarity matrix; A content similarity recall unit (311), used for calculating the content similarity between the source orders and outputting the second recommendation data; A sequence recall unit (313), used to analyze the driver's transportation behavior sequence through a sequence model, predict the driver's transportation demand, and output third recommendation data; The candidate set data is obtained by merging the first recommended data, the second recommended data and the third recommended data.

10. The driver recommendation system based on multimodal feature fusion according to claim 1, characterized in that: Also includes, A model optimization module, connected to the ranking model (24), for training and optimizing the recommendation model using historical data; A recommendation feedback module is connected to the recommendation business layer (3) and is used to obtain feedback data from drivers and update the driver portrait and the recommendation model based on the feedback data.

Images