A Personalized Navigation Route Recommendation Strategy Based on Differential Perceptron Tracking User's Driving Preference

With the increasing frequency of autonomous driving, more and more attention is paid to personalized path planning. However, the path selection preferences of users will change with internal or external factors. Therefore, this paper proposes a personalized path recommendation strategy that can track and study user's path preference. First, we collect the data of the system, establish the relationship with the user preference factor, and get the user's initial preference weight vector by dichotomizing the K-means algorithm. The system then determines whether user preferences change based on a set threshold, and when the user's preference changes, the current preference weight vector can be obtained by redefining the preference factor or calling difference perception. Finally, the road network is quantized separately according to the user preference weight vector, and the optimal path is obtained by using Tabu search algorithm. The simulation results of two scenarios show that the proposed strategy can meet the requirements of autopilot even when user preferences change.


Introduction
In recent years, with the rapid development of information and intelligent transportation systems, artifcial intelligence (AI) has been widely used in transportation tools to provide a comfortable travel experience for drivers and passengers [1][2][3]. Te vehicle Internet tends to share data, enabling vehicles to exchange learning experiences and improve decision-making capabilities. Te individual models are trained on the basis of the data collected. By collecting these learning models from all vehicles, a comprehensive model can be further developed. Tis is of the great signifcance to the application of intelligent transportation system (ITS) in autonomous driving and trafc control.
At present, the existing path navigation and recommendation software in China, such as the Baidu map application and the Josiah Goddard map application, are all popular recommendation software, mainly based on the shortest time [4,5] and the shortest distance [6][7][8], has been unable to meet the growing demand for personalized tourism.
Path recommendation systems usually recommend paths based on the optimization of distance or travel time cost functions. However, the shortest or fastest route is usually not chosen by the driver, so the driver may have diferent recommendations when traveling.
Te personalized route recommendation strategy for each driver's route selection preference has attracted much attention. Te personalized route recommendation strategy not only meets people's personalized needs but also solves the Blythe's paradox, which is widely existed in the current transportation network.
1.1. Literature Review. At present, the driver can get the current driving status of the vehicle through the OBD data. In reference [9], the authors proposed a monitoring system consisting of OBD and GPS, designated area, pricing scheme, and the relationship between other related policies. Te OBD device is used to monitor engine operation and measure fuel consumption and emissions reliably. In reference [10], OBD is used to analyze the driving behavior data through the vehicle preloading equipment and the factors that afect the safe driving and establish the logistic expression model. In reference [11], OBD is used to monitor and prompt hydrocarbon (HC) emissions caused by a failure of the vehicle emission control system. In this paper, the OBD data are used to correlate with the driving user's path selection preferences. Ten, the user's initial preference weight vector is obtained by the bisecting K-means algorithm clustering. In reference [12], Xiong et al. proposed a model predictive control optimal path optimization and tracking framework. First, the relationship between the vehicle and the reference road is established in the road coordinate system, and then, the safe lane under the high constraint environment is established by using the multilayer search method; path boundary and vehicle dynamics constraints are introduced to provide optimal control instructions.
In reference [13], the unsupervised classifcation method based on the bisecting k-means algorithm is applied to the data obtained in low-energy consumption proofread measurement of the online gas analyzer and microsensor. Te good agreement with data from sensors validated the efectiveness of the proposed method. In [14], Zhao and othersproposed a reviewed block-based bag of words model using the bisection K-means clustering method which could signifcantly accelerate the process of codebook generation. In reference [15], a new algorithm based on graph mining is proposed and the bisecting kmeans algorithm is used to fnd frequent terminology collection for the document. Te Tabu search algorithm is a typical shortest path algorithm, which is used to calculate the shortest path from one node to other nodes and is widely used in path planning. In reference [16], the Tabu search algorithm is proposed to solve the problem of Web service composition. It adds a periodic diversifcation step, which is kept on the diversifcation step at the usual completion point. Tis approach combines path relinking with dynamic diversifcation strategies, providing more opportunities for future research. In [17], the Tabu search algorithm is used to improve the route planning strategy in urban areas under trafc congestion, which not only saves driving time but also reduces fuel consumption. In [18], the Tabu search algorithm is used to identify the current topology of the network and help identify the shortest path from the point of failure to the nearest operation source. Te model improves efciency consumption by 23% and bandwidth lifetime by 16%. How to make navigation more efective has been a research hotspot. In [19], Xie et al. proposes a new method of combining global path planning with local path planning, to provide an efcient solution for the unmanned surface vehicle (USV) path planning despite the changeable environment. Te method solves the path planning problem with variable environments and is verifed by simulations and experiments. In [20], Mei proposes an optimal tour guide path planning model based on an ant colony algorithm. Te experimental results show that the proposed model and the optimal path planning algorithm are more optimized. In [21], rough set theory and a genetic algorithm is used in this study to solve the low efcacy and accuracy in robot path planning. Experimental results show that the proposed method can efciently improve robot path planning. In [22], Chen et al. proposed a deep reinforcement learning algorithm for path planning, which has the comprehensive reward function of dynamic obstacle avoidance and goal approaching. Te results show that the method can avoid moving obstacles in the environment, complete the planning task, and has a high success rate.
Te personalized needs of users are not mentioned in these articles. In [23], Long et al. developed a novel route recommendation system to provide real-time personalized route recommendation for self-driving tourists, according to the specifc preferences of users personalized access routes, not only to save the total tour time but also to meet their specifc travel preferences. In [24], a dynamic route guidance method for driver's personalized needs is proposed. Te user preference weight is given artifcially, which has strong subjectivity and lacks objectivity and accuracy, but it provides a reference for the combination of personalized needs and path planning. In reference [25], a personalized path decision algorithm based on user preference is proposed, which lays a good theoretical foundation and framework for this study. But there are also limitations. Te one-time sample trajectory data used in this paper needs a lot of complex mathematical calculations to obtain the data needed for clustering. Te corresponding weight of user driving style obtained by fuzzy c-means clustering is fxed and one-sided, while in real life, the user driving preferences are not always the same and most of them are staged, so it cannot refect the user's personalized driving preferences.
1.2. Contribution. As mentioned above, most traditional path planning schemes improve the speed and precision of path planning algorithmically, while ignoring the efect of user preferences [26][27][28][29][30] on path selection. In scenarios that combine user preferences and path planning, user preferences are unchanged by default. In real life, however, user preferences are not always the same. At diferent times, user preferences may change periodically due to various internal or external factors, such as age, physical condition, climate, and work. Te type of preference may also change over time. At the same time, when the preference type is unchanged, the corresponding preference weights may also change.
Terefore, this paper proposes a personalized path planning strategy that can track and study user preferences. First, the preference type is associated with the OBD data and a threshold is used to select the user's current preference type. Ten, the initial preference weight vector is obtained by the split-k-average algorithm. Finally, the diferential perceptron is used to track and adjust the preference weights in real time, so that the optimal path can be recommended to meet the user's needs when the user's preferences change. In this paper, Nanchang is taken as the reference city of road network quantifcation, and the Tabu search algorithm is used to verify the efectiveness of the strategy.
In the light of the above, the contributions of the present document are summarized as follows: (1) To associate OBD with user preferences and get the initial preference weight vector by clustering (2) Considering the periodic change of preference, the change of two kinds of preference is summarized (3) Tracking the user preferences and fne-tuning the weights through the diferential perceptron (4) Personalized quantization of road network, using the Tabu search algorithm to plan the optimal path (5) Taking Nanchang as a reference city, the simulation experiment is carried out to verify the efectiveness of the proposed strategy

System Implementation Review
A personalized path recommendation system that can track and study the user's path selection preferences is proposed in this paper, as shown in Figure 1. Te system mainly includes data preprocessing module, initial preference module, optimization and adjustment module, and path generation module. When the change of user preferences has been detected, the system will make use of the latest OBD data in the storage space to study independently ofine until a path that can meet the user's current path selection preferences has been found. A brief process of tracking and studying the path recommendation system is shown in Figure 2. In which, three preference factors of time, economy, and comfort are taken as the user's initial path selection preferences, and the path points marked in black are path ofset points. Te frst case is that the preference factors change, the latest recorded OBD data are needed to redetermine them, and the initial driving preference weight is obtained by clustering. In the second case, the diference perceptron needs to be used for learning and correction because the preference factors remain unchanged and the corresponding weight values between the preference factors change. Te third case is that in some road sections, the planned path is not consistent with the actual driving path. Since, the trafc network is real-time and dynamic, emergencies, such as trafc lights, pedestrians, and accidents, will afect the coincidence rate. Tus, in the actual situation, the planning path and the driving path are not exactly consistent. Terefore, within the allowable error range, in this case, it is considered that the user preference remains unchanged and it is considered as a sudden situation.

Data Preprocessing
Module. Diferent self-driving users have diferent path selection preferences, so we conducted a questionnaire on the surrounding self-driving users and obtained the survey data shown in Figure 3. From Figure 3, we can see that four preference factors are afecting the route selection of self-driving users, namely, time, economy, comfort, and safety. Among them, the highest attention is paid to safety. In this paper, we choose these four preference factors as user preference sets.
To determine the type of user preference factors, OBD data and user preference are used to establish a connection in this paper [31]. Te main function of OBD is to supervise the status of components related to emission control during the actual use of vehicles. In this paper, OBD status information, geographic location information, and trip record information are used to correlate user preference factors. Te correlation process is as follows: (1) In the driving process, the greater the average speed of the user, the more the user attaches importance to time. In this paper, v is selected to be the relevance quantity (RQ) of a time-based user and v in any period time can be read directly from the server through the OBD terminal device. (2) Te average fuel consumption (AFC) in the driving process is the most important indicator for economic users. Te lower the value, the higher importance users attach to the economy. Terefore, AFC is chosen as the associated quantity of economic users in this paper. As there is no specifc calculation method for AFC in the standard OBD II and EOBD protocol, it is necessary to use OBD related data for estimation. (3) Te change rate of the relative position of the accelerator pedal (CR ap ) is closely related to the comfortable user. Te lower the (CR ap ) value, the higher the user's attention to the comfort. Users who pay attention to comfort have good driving stability. Tey step on and loose slowly, and give oil smoothly. Terefore, CR ap is chosen as the correlation quantity of comfortable users in this paper. Moreover, it can be obtained directly from the server by modifying the interval time parameters of the standard OBD II and EOBD protocol.
(4) In the process of driving, excessive driving speed will bring a threat to the safety of users. Terefore, the maximum traveling speed of this period is taken as the correlation quantity of safe users in this paper. However, it can be read directly from the OBD terminal device.
Te calculation process of each associated quantity is shown in Table 1.
Considering the laws and regulations related to safe driving and the opinions of professionals in the automobile industry, the threshold value of the average value of daily correlation volume is set, as shown in Table 2.
Te maximum three items of δ i are selected as the user's preference factors by the relevant degree δ i of correlation and threshold calculation. As k � 3, the clustering efect is obvious, and to explain the change of preference factors conveniently, three preference factors are selected as the user's path selection preference each time. Te relevant degree calculation formula is defned as follows:

Initial Habit Module.
After the user preference factor is determined, the initial preference weight vector is obtained by clustering. Time, economy, and comfort are taken as examples and the bisecting k-means algorithm is used in clustering in this paper. To facilitate the clustering and efect display of OBD data, the data should be normalized before clustering. Te partial results of some user OBD sampling data processed are shown in Table 3. v, AFC, and CR ap are used to establish the three-dimensional coordinate system, and then, the normalized data are divided into the bisecting K-means clustering, and the initial preference weight vector of users is obtained according to the clustering results. In order to make the clustering efect signifcant, here, select Δt � 1 min to be clustered, and the results are shown in Figure 4.

Computational Intelligence and Neuroscience
As can be seen from the fgure, the value of red marked points v-axis is generally higher than the other two categories. Due to the highest attention to time, the value v of time users is generally large. Terefore, the red marker class is a time-based feature point and Te values of the AFC-axis of the points marked in green are generally smaller than those of the other two categories. Economic users pay the highest attention to fuel consumption, and the value of AFC is generally very small. Terefore, the green marker is the economic characteristic point. Te value of CR ap -axis of blue marked points is generally smaller than the other two categories. Comfortable users pay the most attention to driving stability and the value of CR ap is generally very small. Terefore, the blue marker is the comfort feature point. After the bisecting K-means clustering converges, the number of coordinate points belonging to three clustering centers is normalized to get the initial user preference weight vector w � (w 1 , w 2 , w 3 ). Te normalization formula is as follows: Among them, n 1 , n 2 , and n 3 are the numbers of coordinate points belonging to the three clustering centers individually.

Path Generation Module.
Tis part is mainly divided into two parts. Te frst part describes how to individualize and quantify the road network to establish a model for solving the optimal path. Te second part shows how to calculate the optimal path by simulation.

Road Network Quantifcation.
To simplify the road model and calculation, referring to the regulations of highway technical engineering standard, urban road design code, urban road network planning index system, Nanchang urban trafc planning, and combining with the actual characteristics of the road network, Baidu map and the results of feld investigation and verifcation, the roads in Nanchang are divided into the following fve categories. To sum up, the road conditions in Nanchang are actual. Te actual situation of each road is shown in Table 4.
In order to unify the quantitative standard and simplify the calculation of the optimal path with a genetic algorithm, the data in Table 3 are normalized and correlated. Te process is as follows: Cos t E � EC, Cos t S � 0.5 * MS 500 + 0.5 * RS.
Te data processed by formula (3) are shown in Table 5.
Based on the comprehensive consideration of diferent driving preferences and simplifed calculation, the unit cost function is designed as follows: While car running 3: Read and storage the values of v, CR ap and V max every 6 seconds from OBD data 4: Read the values of, L, R PM , M AP , and I AT every 6 seconds from OBD data 5: Calculate Computational Intelligence and Neuroscience In this paper, each user has three main driving preferences, that is, the other three preference weights are set to zero, making m � 3, so the cost of personalized quantifcation of the K segment road is as follows:

Tabu Search
Algorithm. After the recommendation system personalizes and quantifes the road network according to the user preference weight vector [32], the model becomes a classic problem of fnding the optimal solution. In this paper, the Tabu search algorithm is used to calculate the optimal path according to the user's preference. Here, the model diagram shown in Figure 5 is selected to illustrate how the algorithm calculates the optimal path. In Figure 5, the cost consumption values of each point between 1 and 20 are calculated, respectively, by formula (5). After personalized quantifcation of the road network, fnding the optimal path becomes a single source shortest path problem shown in Figure 5. Te process of using the Tabu search algorithm to get the optimal path of the model shown in Figure 5 is shown in Table 6. As can be seen from Table 6, the fnal optimal path is 1 ⟶ 4 ⟶ 7 ⟶ 10 ⟶ 13 ⟶ 15 ⟶ 16 ⟶ 20, and the total cost consumption is 88.

Optimization and Adjustment Module.
In order not to afect the user's self-driving experience, it is stipulated that the system will conduct tracking and studying at night every day, and the preference tracking process is shown in Table 7. Due to the existence of emergencies, the consistency between the planned path and the actual driving path will not always be 100%. In the system, a threshold parameter is set for the coincidence degree. Before training every day, the system will flter the OBD data of the day and eliminate the data whose coincidence degree is higher than the threshold value. Ten, the system recalculates the correlation degree δ i of four correlation quantities according to the remaining OBD data of the day in the storage space and sorts them from small to large. If the correlation quantity corresponding to the minimum value    Figure 4: Bisecting K-means clustering results (above is before clustering and below is after clustering).    Computational Intelligence and Neuroscience of δ i changes, the system determines that the preference factor changes, and clusters the latest data of the other three correlation quantities to get the initial weight vector again. On the contrary, the system judges that the preference factor has not changed, and calls the diferential perceptron to fne-tune the weight vector. Because of the relatively small number of data samples on the same day, direct clustering by ignoring the previous data will enlarge the efect of change. Data clustering together will cover up smaller weight changes. Terefore, in this paper, we choose to call the diferential perceptron to fne-tune to improve the accuracy. Te setting of the diferential perceptron function is shown in Table 8. Taking OBD data and current weight vector as input, the maximum number of iterations is set as 2,000. Te total length of fve types of roads is calculated for the actual path and the planned path, respectively, and the total cost corresponding to the three preference factors is calculated for the two cases. Ten, the change of the weight of each preference factor is calculated. Among which, η is the learning efciency, take 0.003 to obtain the weight vector closest to the current user path selection preference through repeated iterative learning.

Experiment
Two types of preference change are proposed in this paper: one is the change of user preference factor and the other is the change of the weight of the same preference factor. To verify the efectiveness of the strategy proposed in this paper, Nanchang city is taken as a reference City, and simulation experiments are carried out in two scenarios, respectively. In scenario 1, the user preference factor changes. Assuming that the initial preference factor of user A is economy, comfort, and safety, then due to work reasons, time is more important than the economy, so the preference factor becomeseconomy, comfort, and safety. In scenario 2, the weight of the same user preference factor changes. Suppose that the initial preference factor of user B is time, economy, and comfort. Ten, due to the fnancial crisis of the family, user B pays more attention to the economy and less attention to comfort. In these cases, 90% of the threshold value is used to determine whether it is an emergency or not.

Experiment 1. Te preference factor of self-driving user
A is economy, comfort, and safety, and the corresponding initial weight vector is obtained by the clustering algorithm, which is w � (0.48, 0.14, 0.38). After the preference factor of user A changes, the planned path and the actual driving path are shown in the red path and green path in Figure 6, respectively. It can be seen from the fgure that due to the change of preference factors, the coincidence of the planning path and the actual path are very low, only 0.785%, indicating that the current planning strategy can no longer meet the personalized traveling needs of users.
After the personalized quantifcation of the road network w � (0.48, 0.14, 0.38), the specifc generation values of economic consumption (TC), comfort consumption (EC), and safety consumption (CC) in two cases are shown in Table 9, respectively.
Te total cost C total of the actual path and the planned path can be obtained through classifcation and integration of the data of the road type in Table 9 and is shown in Table 10. As can be seen from Table 10, due to the change of preference type, the path types in the two cases are diferent, and the total cost is also very diferent, which is 12.737.
When the system is tracking and studying at night, the correlation δ i of four preference factors is calculated by formula (1). Among them, the correlation quantity of Calculate δ i with the latest OBD data, i ∈ T, E, C, S 3: If RQ of δ i min unchanged 4: w′ � Improved diferential perceptron (the last OBD data, w) 5: Else 6: Obtain w′ by reclustering the OBD data of the other three RQ (7) End if (8) End Table 8: Te improved diferential perceptron algorithm. Algorithm 3: Improved diferential perceptron Input: OBD data, habit weight vector Output: Improved new habit weight vector w′ � w 1 w 2 w 3 11: While n ! � 2000 or ‖w′ − w‖ > ε 12: For each road classifcation do 13: Calculate the total length of the section in AP and PP respectively 14:

15:
End for 16: For each habit factor do

17:
Calculate the total cost in AP and PP respectively 18:

19:
Calculate weight change volume respectively 20: ∆ω � η( U APi − U PPi )i ∈ T, E, C, S 21: Calculate and normalize the new driving style weight 22: ω i j � ω j + Δω j � 1, 2, 3 23: End for 24: n � n + 1 25: Re-quantify road network with w ′ 26: Replan the optimal path with the Tabu search algorithms 27: End while 28: Return w′ � w 1 w 2 w 3 8 Computational Intelligence and Neuroscience minimum δ i changes from the previous time to economy. Terefore, the system determines that the user A's preference factor changes, and the current preference factor changes to time, comfort, and safety. Te current weight vector is obtained by the bisecting K-means clustering. Finally, the planning path calculated by the Tabu search algorithm using personalized quantitative road network is shown in the red path in Figure 7.
As can be seen from Figure 7, after tracking and studying, the consistency between the planned path and the actual path is greatly improved, from 0.785% to 98%. Among them, the specifc cost data in the two cases after adjustment are shown in Table 11.
Te total cost C total of the actual path and the planned path can be obtained through classifcation and integration of the data of the road type in Table 11 and is shown in     Table 12 As can be seen from Table 12, the path types in the two cases are the same, and the total cost diference is only 0.081, which indicates that the system can still meet the user's path planning needs through tracking and adjusting after the user's preference factor changes.

Experiment 2.
Te preference factor of self-driving user B is time, economy, and comfort, and the corresponding initial weight vector is w � (0.14, 0.44, 0.42). After user B's preference weight changes, the planned path and the actual driving path are shown in red and green paths in Figure 8, respectively. Due to the change of preference weight, the consistency between the planned path and the actual path is not high, which is 76.64%. It shows that the current planning strategy does not fully meet the user's personalized travel needs. To improve the degree of coincidence, we need to further improve the accuracy of the weight vector.
After the personalized quantifcation of the road network w � (0.14, 0.44, 0.42), the specifc generation values of time consumption (TC), economic consumption (TC), and comfort consumption (EC) in two cases are shown in Table 13, respectively.
Te total cost C total of the actual path and the planned path can be obtained through classifcation and integration of the data of the road type in Table 13 and is shown in Table 14.
It can be seen from the table that, due to the change of preference weight, the path types in the two cases are the same, but the length is diferent, and the total cost diference is 2.149.
Similarly, when the system is tracking and studying at night, the correlation δ i of four preference factors is recalculated by formula (1). Among them, the correlation quantity of the corresponding minimum δ i value does not change, which is safety. Terefore, the system will determine that the user B preference type has not changed but the weight changes, and call the diferential perceptron to fnetune, and fnally, get the current weight vector w ′ � (0.15, 0.51, 0.34). Finally, the planning path calculated by the Tabu search algorithm using personalized quantitative road network w ′ is shown in the red path in Figure 9.
As can be seen from Figure 9, after tracking and studying, the consistency between the planned path and the actual path has improved, from 76.64% to 84.58%. Among them, the specifc cost data in the two cases after adjustment are shown in Table 15.        Figure 9: Te initial planning path and the actual driving path before adjustment (the left one is the model diagram and the right one is the actual map). Te total cost C total of the actual path and the planned path can be obtained through classifcation and integration of the data of the road type in Table 15 and is shown in Table 16. As can be seen from Table 16, the path types in the two cases are the same, and the total cost diference is 1.199. Compared with studying before adjustment, the diference in total cost decreases and the coincidence increases. It shows that the system can also meet the user's path planning needs by tracking and adjusting when the user's preference weight changes.

Conclusion and Outlook
A personalized path recommendation strategy that can track and study the user's path selection preferences is proposed in this paper. In the case that the self-driving users do not know their preferences and their preferences will change periodically, the system frst determines the main driving preference factors of users through the collected OBD data and obtains the initial preference weight vector through the bisecting K-means clustering. When the preference changes, the system judges according to the relevant parameters and adjusts the preference weight vector according to the situation. Ten, the road network model is quantifed according to the weight vector. Finally, the Tabu search algorithm is used to calculate the optimal path. In the two scenarios, through the comparison between the planned path and the actual path and the path before and after tracking and studying, the efectiveness of the strategy proposed in this paper is verifed. Te results show the consistency between the planned path and the actual path has improved from 76.64% to 84.58% after tracking and studying the user's personalized driving behavior. Te adjusted total cost of initially planned route is 18.580, while actual path 19.779. Te path types of the two cases are the same, and the total cost diference is only 0.081. It is proved that the strategy can also meet the personalized traveling needs of users when their preferences change.
To make the Tabu search algorithm not easily fall into local optimum when calculating the optimal path, the road network model in this paper is relatively simple. However, when the actual road network is larger and the route is more complex, the algorithm is easy to fall into local optimum. At the same time, only static factors rather than the real-time trafc situations in the road network are considered in this paper. In the future work, we should improve the accuracy of planning path algorithm or change the way of road network modeling, and add some dynamic factors to make the strategy proposed in this paper more universal.

Data Availability
Te data used to support the fndings of this study are available from the corresponding author upon reasonable request (pzchen@tzc.edu.cn).

Conflicts of Interest
Te authors declare that they have no conficts of interest.