A method for predicting instant delivery routes based on multi-source data
Through the deep neural network model of multi-source data fusion, the rider data characteristics are extracted and fused, and the instant delivery path is predicted, which solves the problem of insufficient single-source data and improves the path prediction accuracy and delivery efficiency.
Patent Information
- Application Number
- CN202111301713.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-11-04
- Publication Date
- 2024-12-06
- Estimated Expiration
- 2041-11-04
AI Technical Summary
The existing path prediction algorithm is based on single source data and cannot effectively deal with multiple complex factors in real-time delivery, resulting in low accuracy of rider path prediction, affecting order distribution and delivery efficiency.
Using a deep neural network model with multi-source data fusion, we use ST deep neural network to predict the rider's delivery path by extracting and fusion of features in rider dots, encounters and GPS data.
It improves the accuracy of path prediction, reduces the impact of the quality of a single data source, enhances the model learning ability and training efficiency, and improves the distribution efficiency.
Smart Images

Figure CN114004409B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to an instant delivery route prediction method based on multi-source data, and belongs to the technical field of computer data processing. Background Art
[0002] To ensure service quality, instant delivery services have strict order delivery deadlines. However, due to factors such as traffic, weather, and time and space, order delays are inevitable. Accurately predicting delivery routes for instant delivery drivers helps assign new orders to the most suitable drivers, improve delivery efficiency, and reduce order delays.
[0003] Most existing route prediction algorithms are based on single-source data, such as GPS trajectory data or order logistics information. These methods cannot cope with the various complex factors in real-time delivery, such as multiple destinations and individual differences among delivery drivers. In addition, most existing research tends to use heuristic algorithms to optimize delivery routes for delivery drivers and recommend the optimized delivery routes to the delivery drivers. These studies assume that the actual delivery routes of delivery drivers are consistent with the recommended routes. However, in reality, delivery drivers may prefer to plan routes based on their own experience rather than strictly follow recommendations. According to research, at least 20% of delivery drivers' decisions differ from the recommended routes. Such problems lead to a decline in the performance of order dispatch algorithms.
[0004] It can be seen from this that existing technologies are unable to achieve accurate rider path prediction. Order dispatch based on the rider path prediction results with low accuracy has low delivery efficiency, which affects the service quality of instant delivery services. Therefore, a new solution is urgently needed to solve the above technical problems. Summary of the Invention
[0005] The present invention aims to solve the problems existing in the existing technology and provides an instant delivery route prediction method based on multi-source data. This technical solution uses a deep neural network model to accurately predict the rider's delivery route by fusing multi-source heterogeneous data.
[0006] To achieve the above objectives, the technical solution of the present invention is as follows: an instant delivery route prediction method based on multi-source data, characterized by comprising a prediction system, wherein the prediction system includes a multi-source feature extraction module to achieve data fusion at the feature level; and a route prediction module based on a deep neural network, which predicts the rider's route based on the multi-source features, completing feature fusion at the model level;
[0007] The method comprises the following steps:
[0008] S1. Extract rider reporting data, rider encounter data, and rider GPS data from the instant delivery platform;
[0009] S2. Considering the independence and correlation of multi-source data, multi-source feature extraction and fusion are performed on multi-source data;
[0010] S3. Use feature conversion layers to extract high-dimensional feature representations of multi-source features and predict instant delivery routes based on deep neural networks.
[0011] As an improvement of the present invention, S1. extracts rider reporting data, rider encounter data, and rider GPS data from the instant delivery platform, specifically as follows: the platform requires riders to actively report order progress on the mobile application when accepting orders, arriving at merchants, picking up orders, and delivering orders. The platform obtains the real-time status of orders and riders through the riders' reporting data; the platform uses Beacon technology to collect rider encounter information, and its principle is as follows: through the mobile application, the rider's mobile phone continuously broadcasts an ID tupes through the BLE 5.0 (Bluetooth 5.0) protocol during its working hours and continuously scans to sense Bluetooth signals, and then uploads the scanned ID tupes to the back-end server in real time through the network connection. After the server receives the rider ID and the ID tupes received by the rider, it extracts the rider encounter data; the platform obtains the rider GPS data through the GPS positioning of the rider's mobile phone.
[0012] As an improvement of the present invention, the multi-source feature extraction and fusion work in step S2 extracts the multi-source combined features of the target rider's destination and the target rider's meeting rider's destination from the multi-source data, realizing data fusion at the feature level. The specific work is as follows:
[0013] The three types of features extracted from multi-source data are: encounter features, GPS features, and incentive features. For each destination of the target rider, three types of features are extracted; for each destination of the target rider who was recently encountered during the delivery, two types of features (incentive features and encounter features) are extracted, and then the Cartesian product of the destination features of these two riders (target rider and encounter rider) is taken to obtain the combined features, that is, A×B={(a, b)|a∈A∧b∈B}. Each combined feature represents a combination of the target rider's destination and the encounter rider's destination; specifically, when the target rider has N destinations and the rider he encounters during the delivery has M destinations, the model first extracts the target rider's N destination features and the encounter rider's M destination features from the rider's reporting data, encounter data, and GPS data, and then obtains M×N combined features.
[0014] As an improvement of the present invention, the three types of features are specifically defined as follows:
[0015] A. Encounter features: Encounters occur on the way from a certain starting point to a destination. By correlating rider encounter data with rider point reporting data, we can extract encounter features such as the rider’s starting point location when the encounter occurs. Table 2 shows the definition of encounter features. Assume that the target rider c p and met rider c e The next destinations are d i and d j , if according to the target rider c p and met rider c e Encounter characteristics x E ={L o , L d , t e} Finding that they cannot meet, we infer that the next destination is d i and d j The assumption is wrong, therefore, the encounter characteristics help to verify the prediction of the next destination;
[0016] B.GPS features, assuming the next destination is d i , then the GPS features of the rider at time t include the geographic location L of the rider’s current location extracted from the GPS data g , and L extracted from the dot reporting data o , L d , using GPS feature x G ={L g , L o , L d}, determine whether the rider’s current trajectory direction is consistent with the direction to reach the destination d i Consistent, thus verifying the destination d i Is it the next destination?
[0017] C. Incentive features, for rider c, at time t, he has a set of destinations D c,t , D c,t is the destination set of rider c at time t. The size of the destination set is N, and the incentive features of these destinations are extracted from the reporting data. The motivational features are mainly composed of three types of features: time, space, and scene. Among them, the rider c, the target rider for path prediction is c p , the target rider's encounter rider is c e .
[0018] As an improvement to the present invention, the ST (Specialized Transition) deep neural network designed in step S3 completes the feature fusion at the model level, thereby realizing the next destination prediction of the target rider. The specific work is as follows:
[0019] The key part of the route prediction module is the ST network based on deep neural network. The ST network predicts the possibility of each rider destination combination. Its structure is as follows Figure 2 As shown. ST network for stimulus feature x I , Encounter Features x E and GPS feature x G The three types of features undergo different feature transformations. Each feature transformation consists of three layers: input layer, hidden layer, and output layer. The new feature vectors obtained after feature transformation are finally connected together through a fully connected layer, and the sigmoid activation function is used to output the score.
[0020] The new feature vector k obtained by feature transformation is k=K(x|W h , W o , b h , b o ) represents, where K(·) is the h , W o , b h , b o The parameterized feature conversion function, K(·) is decomposed into
[0021] h=ReLU(W h x+b h ),
[0022] k=ReLU(W o h+b o );
[0023] Among them, x is the input feature vector, h is the output of the hidden layer, and W h , b h is the weight and bias of the hidden layer, W o , b o are the weights and biases of the output layer. The activation function used in the hidden layer and the output layer is ReLU. The parameters of different feature conversion functions are independent of each other.
[0024] For the target rider's destination d i , the new GPS vector k is obtained by GPS feature conversion G ,
[0025]
[0026] GPS feature with riders Compared with d i There is a stronger correlation with the likelihood of being selected as the next destination,
[0027] Likewise, destination-based incentive features New stimulus vector Respectively represent the destination di , d j Rider c p , c e the possibility of choice;
[0028]
[0029]
[0030] One of the characteristics of the ST network is the shared target rider c p and met rider c e The design reduces the number of neurons required and effectively saves training costs.
[0031] Because it is necessary to jointly analyze the encounter characteristics of the target rider and the encounter rider to obtain the possibility of the rider encounter, it is necessary to combine the two riders c p , c e Encounter characteristics Splicing, and then input to the encounter feature conversion layer to obtain the new encounter vector k E
[0032]
[0033] The new vector k E is closely related to the probability of encounter, which in turn is closely related to the correctness of the next destination,
[0034] Finally, the ST network jointly analyzes these new vectors and outputs the target rider destination d i The final prediction score p of
[0035]
[0036] Where W f , b f are the weights and biases of the fully connected layer;
[0037] The prediction task of the ST network is to determine whether the destination combination is correct. Let y be the sample label, with a value of 0 or 1. The network is trained using the cross entropy loss function, that is, the loss function is
[0038] Loss=-[yln(p)+(ly)ln(1-p)],
[0039] For each combination of features, the prediction score is obtained through the ST network, and the scores are sorted. The destination of the target rider represented by the combination of features with the highest score is the prediction result (i.e., the next destination of the target rider);
[0040]
[0041] p=M ST (x cb ),
[0042]
[0043]
[0044] The entire path prediction process can be expressed by the above equation, represents the predicted next destination of the target rider, M ST (·) represents the ST network function, F toL (·) indicates a combination of features The mapping function from the target rider destination d.
[0045] Compared with the existing technology, the present invention has the following advantages: 1) The present invention is the first to predict the rider path of instant delivery based on multi-source data. The prediction method based on multi-source data not only provides fine-grained spatiotemporal information, but also helps to eliminate the impact of low data quality from a single data source. This method can effectively improve the prediction accuracy. 2) The feature conversion layer proposed in the present invention extracts effective information from multi-source features, reduces the complexity of the model, enhances the learning ability of the model, accelerates the convergence speed of the model, and saves model training time. BRIEF DESCRIPTION OF THE DRAWINGS
[0046] Figure 1 It is a schematic diagram of the overall framework of the present invention;
[0047] Figure 2 Schematic diagram of the ST deep neural network structure of the present invention. DETAILED DESCRIPTION
[0048] In order to deepen the understanding of the present invention, this embodiment is described in detail below with reference to the accompanying drawings.
[0049] Example 1: See Figure 1 、 Figure 2 , an instant delivery route prediction method based on multi-source data, e.g. Figure 1 The figure shows the technical framework of the present invention. The instant delivery route prediction technology based on multi-source data proposed by the present invention includes the following steps:
[0050] S1. Extract rider reporting data, rider encounter data, and rider GPS data from the instant delivery platform. This specifically involves the following tasks.
[0051] The platform requires riders to proactively report order progress on their mobile app when accepting an order, arriving at the merchant, picking up the order, and delivering the order. The platform obtains the real-time status of the order and the rider through the rider's reporting data.
[0052] The platform uses Beacon technology to collect rider encounter information. The principle is as follows: through the mobile phone application, the rider's mobile phone continuously broadcasts an ID tupes through the BLE 5.0 (Bluetooth 5.0) protocol during his working hours and continuously scans to sense Bluetooth signals. The ID tupes received by the scan are then uploaded to the back-end server in real time through the network connection. After the server receives the rider ID and the ID tupes received by the rider, it extracts the rider encounter data;
[0053] The platform obtains the rider’s GPS data through the rider’s mobile phone GPS positioning.
[0054] S2. Considering the independence and correlation of multi-source data, perform multi-source feature extraction and fusion on multi-source data. This includes the following tasks:
[0055] N destination features of the target rider and M destination features of the encountered rider are extracted from the rider’s reporting data, encounter data, and GPS data. For each destination of the target rider, three types of features are extracted (encounter features, GPS features, and incentive features). And for each destination of the encounter rider that the target rider has recently encountered, two types of features (incentive features and encounter features) are extracted. Then the Cartesian product of the destination features of the two riders (target rider and encounter rider) is taken to obtain M×N combined features, that is, A×B={(a, b)|a∈A∧b∈B}. Each combined feature represents a combination of the target rider’s destination and the encounter rider’s destination.
[0056] The three types of features are extracted as follows:
[0057] A. Encounter Features. The rider encounter data is correlated with the rider's point reporting data to extract encounter features such as the rider's starting location at the time of the encounter. See Table 2 for the specific definition of encounter features.
[0058] B. GPS features. Extract the rider’s current location L from the GPS data g , extract the geographic location L of the starting point and destination from the reported data o , L d , get GPS feature x G ={L g , L o , L d},in:
[0059] L is the geographical location of the location. L = (x, y), where x and y represent the longitude and latitude coordinates of the location, respectively.
[0060] Table 2 Encounter characteristics of the present invention
[0061]
[0062] C. Incentive features. For rider c, the destination D at time t c,t , extract set D from the reporting data c,t Incentive characteristics of various destinations in The motivational features are mainly composed of three types of features: time, space, and scene. The specific definitions are shown in Table 3.
[0063] Table 3 Excitation characteristics of the present invention
[0064]
[0065]
[0066] Where: E(·), E(x, y) are functions for determining equality. When x=y, E(x, y)=1, otherwise E(x, y)=0.
[0067] S3. Use the feature conversion layer to extract high-dimensional features from multiple sources and predict instant delivery routes based on a deep neural network. This includes the following tasks:
[0068] Construct a deep neural network ST, whose structure is as follows Figure 2 As shown. The ST network stimulates the feature x in the combined feature I , Encounter Features x E and GPS feature x G The three types of features are transformed differently. K(·) is the h , W o , b h , b o The parameterized feature conversion function, K(·) is decomposed into
[0069] h=ReLU(W h x+b h ),
[0070] k=ReLU(W o h+b o ).
[0071] Among them, x is the input feature vector, h is the output of the hidden layer, and W h , b h is the weight and bias of the hidden layer, W o , b o are the weights and biases of the output layer. The activation function used in the hidden and output layers is ReLU. The parameters of different feature conversion functions are independent of each other.
[0072] The GPS characteristics of the target rider are converted into Convert to get the new GPS vector k G ,
[0073]
[0074] Destination incentive features are converted into incentive features through incentive feature conversion Convert to a new stimulus vector
[0075]
[0076]
[0077] By placing the two riders c p , c e Encounter characteristics Splicing, input to the encounter feature conversion layer to obtain the new encounter vector k E ,
[0078]
[0079] Jointly analyze new vectors through fully connected layers Output target rider destination d i The final prediction score p of
[0080]
[0081] Where W f , b f are the weights and biases of the fully connected layer.
[0082] The ST network is trained using the cross entropy loss function. The loss function formula is:
[0083] Loss=-[yln(p)+(ly)ln(1-p)].
[0084] For each combination of features, a prediction score is obtained through the trained ST network. The scores are sorted, and the destination of the target rider represented by the combination of features with the highest score is output as the prediction result (i.e., the target rider's next destination). It can be expressed as follows:
[0085]
[0086] p=M ST (x cb ),
[0087]
[0088]
[0089] The next destination of the target rider predicted by Youshi, M ST (·) represents the ST network function, F toL (·) indicates a combination of features The mapping function from the target rider destination d.
[0090] It should be noted that the above embodiments are not intended to limit the scope of protection of the present invention, and equivalent changes or substitutions made on the basis of the above technical solutions fall within the scope of protection of the claims of the present invention.
Claims
1. A method for predicting instant delivery routes based on multi-source data, characterized in that: It includes a prediction system, which includes a multi-source feature extraction module to achieve data fusion at the feature level; and a path prediction module based on a deep neural network to predict the rider's path based on multi-source features and complete feature fusion at the model level; The method comprises the following steps: S1. Extract rider reporting data, rider encounter data, and rider GPS data from the instant delivery platform; S2. Consider the independence and correlation of multi-source data, and extract and fuse multi-source features of multi-source data; S3. Use feature conversion layer to extract high-dimensional feature expression of multi-source features and predict instant delivery path based on deep neural network; Among them, the multi-source feature extraction and fusion work in step S2 extracts the multi-source combination features of the target rider's destination and the target rider's encounter rider's destination from the multi-source data, realizing data fusion at the feature level. The specific work is as follows: The three types of features extracted from multi-source data are: encounter features, GPS features and incentive features. For each destination of the target rider, three types of features are extracted; for each destination of the target rider who was recently encountered during the delivery, two types of features, namely incentive features and encounter features, are extracted, and then the Cartesian product of the destination features of these two riders, namely the target rider and the encounter rider, is taken to obtain the combined features, namely A×B={(a, b)|a∈A∧b∈B}, and each combined feature represents a combination of the target rider's destination and the encounter rider's destination; specifically, when the target rider has N destinations, and the rider he encounters during the delivery has M destinations, the model first extracts the target rider's N destination features and the encounter rider's M destination features from the rider's reporting data, encounter data and GPS data, and then obtains M×N combined features; The ST (Specialized Transition) deep neural network designed in step S3 completes the feature fusion at the model level and realizes the prediction of the next destination of the target rider. The specific work is as follows: The key part of the route prediction module is the ST network based on deep neural network. The ST network predicts the possibility of each rider destination combination. The ST network is based on the incentive feature x I , Encounter Features x E and GPS feature x G The three types of features are transformed differently. Each feature transformation consists of three layers: input layer, hidden layer and output layer. The new feature vectors obtained after feature transformation are finally connected together through the fully connected layer, and the sigmoid activation function is used to output the score. The new feature vector k obtained after feature transformation is k=K(x|W h , W o , b h , b o ), where K(·) is the h , W o , b h , b o The parameterized feature conversion function, K(·) is decomposed into h=ReLU(W h x+b h ), k=ReLU(W o h+b o ); Among them, x is the input feature vector, h is the output of the hidden layer, and W h , b h is the weight and bias of the hidden layer, W o , b o are the weights and biases of the output layer. The activation function used in the hidden layer and the output layer is ReLU. The parameters of different feature conversion functions are independent of each other. For the target rider's destination i , the new GPS vector k is obtained by GPS feature conversion G , GPS features with riders Compared with d i The likelihood of being selected as the next destination has a stronger correlation based on the destination motivational characteristics New stimulus vector Respectively represent the destination d i , d j C p , c e Possibility of choice; One of the characteristics of the ST network is the shared target rider c p and meet rider c e Parameters of the excitation feature conversion layer; Because it is necessary to jointly analyze the encounter characteristics of the target rider and the encounter rider to obtain the possibility of the rider encounter, it is necessary to combine the two riders c p ,c e Encounter characteristics concatenate, and then input into the encounter feature conversion layer to obtain the new encounter vector k E The new vector k E is closely related to the probability of encounter, which in turn is closely related to the correctness of the next destination. Finally, the ST network jointly analyzes these new vectors and outputs the target rider destination d i The final prediction score p of Where W f , b f are the weights and biases of the fully connected layer, The prediction task of the ST network is to determine whether the destination combination is correct. Let y be the sample label, with a value of 0 or 1. The cross entropy loss function is used to train the network, that is, the loss function is Loss=-[yln(p)+(ly)ln(1-p)], For each combination of features, the prediction score is obtained through the ST network, and the scores are sorted. The destination of the target rider represented by the combination of features with the highest score is the prediction result (i.e., the next destination of the target rider); p=M ST (x cb ), The entire path prediction process can be expressed by the above equation, represents the predicted next destination of the target rider, M ST (·) represents the ST network function, F toL (·) indicates a combination feature The mapping function to the target rider destination d.
2. The instant delivery route prediction method based on multi-source data according to claim 1 is characterized in that: S1. Extract rider reporting data, rider encounter data, and rider GPS data from the instant delivery platform, as follows: the platform requires riders to actively report order progress on the mobile application when accepting orders, arriving at merchants, picking up orders, and delivering orders. The platform obtains the real-time status of orders and riders through the riders' reporting data; the platform uses Beacon technology to collect rider encounter information, and the principle is as follows: through the mobile application, the rider's mobile phone continuously broadcasts an ID tupes through the BLE 5.0, i.e. Bluetooth 5.0 protocol during its working hours and continuously scans to sense Bluetooth signals, and then uploads the scanned ID tupes to the back-end server in real time through the network connection. After the server receives the rider ID and the ID tupes received by the rider, it extracts the rider encounter data; the platform obtains the rider GPS data through the GPS positioning of the rider's mobile phone.
3. The instant delivery route prediction method based on multi-source data according to claim 2 is characterized in that: The three types of features are defined as follows: A. Encounter feature: The encounter occurs on the way from a certain starting point to the destination. By associating the rider encounter data with the rider’s reporting data, the encounter feature of the rider’s starting point when the encounter event occurs is extracted. Assuming that the target rider c p and meet rider c e The next destinations are d i and d j , if according to the target rider c p and meet rider c e Encounter characteristics x E ={L o , L d , t e } Finding that they cannot meet, we infer that the next destination is d i and d j The assumption is wrong, B. GPS feature, set the next destination to d i , then the rider’s GPS features at time t include the rider’s current location extracted from the GPS data, g , and L extracted from the dot reporting data o , L d , using GPS feature x G ={L g , L o , L d }, determine whether the rider’s current trajectory direction is consistent with the direction to the destination d i Consistent, thus verifying the destination d i Is it the next destination? C. Incentive features, for rider c, at time t, he has a set of destinations D c,t , extract the incentive features of these destinations from the reporting data The motivational features are composed of three types of features: time, space, and scene.
