About The Theme-Based Research Scheme
The Theme-Based Research Scheme is one of the important projects of the Drip GAIA Research Cooperation Program. Based on the concept of open collaboration, the project realizes the transformation from scientific research innovation to industrial value and promotes innovation in the industry through the organic combination of cutting-edge academic research and actual business scenarios, Through the opening of business scenarios, Didi hopes to find and define problems with the academic community, work together to solve problems in the field, build high-level cross-border knowledge and research networks, and build a community of industry-university-research cooperation.
The 2018 Fall Thematic Research Scheme includes 15 real-life business scenarios from DiDi, including machine learning, computer vision, speech signal processing, geographic information technology, and energy and automotive research. The world's top scholars and Didi Researchers are invited to discuss the application of cutting-edge technology in the field of travel and drive technological innovation with real-life scenarios.
The DiDi Theme-Based Research Scheme is divided into two application seasons each year: spring and autumn. The deadline for submitting applications in the fall of 2018 is August 19th. Scholars from global academic institutions or non-profit laboratories are invited. Applicants can submit project proposals through the official website or email. The detailed application process and guidelines are detailed in the “DIDI Research Outreach Initiative” on WeChat and the official website.
The official website:
https://outreach.didichuxing.com/RFP
Email of GAIA Research Cooperation Program:
Next, we will deeply interpret the following topics: big data-based optimization research of electric vehicle powertrain system, geographic information technology, voice technology, computer vision and machine learning in the travel scene. We hope that you can find a report that matches your research direction.
Topic one：Research on Optimization of Electric Vehicle Powertrain System Based on Big Data
1.Research Introduction
The current algorithms and calibrations of EV powertrain are based on limited experiments and experience, which could not make real-time adjustments according to the changing environment, driving habits, vehicle condition and other factors in the life cycle of the vehicle. If the strategy is either too conservative or too aggressive, the vehicle’s maximum operation capability and safety cannot be guaranteed.
Therefore, in order for the vehicle to operate in the best condition and to maximize value, it is necessary to combine the real-time data with historical big data to optimize the control algorithms and calibration parameters.
2.Target
By mining the characteristics featuring the vehicle's driving conditions, including speed, mileage, temperature, charging and other information, an environment in which power system simulation and development based on big data is provided.
1.SOH of the battery system is expected to be estimated by collecting the current, voltage, temperature and other parameters of the cell in real time.
2.Based on the historical data of the vehicle, predictions on the dynamic and economic changes of the vehicle and the remaining mileage of the battery system is demanded.
3.On the premise of ensuring that the SOH of the battery is not damaged and the driving performance of the whole vehicle is within an acceptable range, algorithm or parameters such as cooling strategy, energy recovery strategy and discharge strategy are expecte to extend the range by one time charging and ranges in the vehicle's whole life cycle.
Topic Two：Geographic Information System
1. Route Planning for Smart Transportation
1.1 Background
Route planning plays a fundamental role in many DiDi higher-tier business operations and is also an important basic service in the field of smart transportation, affecting the travels of hundreds of billions of people. How to make the best use of static road network attribute information along with dynamic road traffic condition information to provide the user with an optimal route is a challenging task within the industry and a valuable research topic.
1.2 Research Target
For the above key issues, research objectives include but are not limited to the following:
1） Route Quality Evaluation Method
Considering the tremendous amount of user trajectories, the basic responsibility of the platform lies in evaluating the quality of the outputs of the route planning algorithm and the extent of their acceptance by drivers and passengers, which furthermore clarifies the direction for further optimization. An effective and automatic quality evaluation method involves multiple dimensions, including route length, time-cost, price, traffic condition, user preferences, road levels, etc. Different business scenarios, such as hitching and pooling also warrant additional consideration.
2） Road Network Weights Mining
Common route planning algorithms usually abstract the actual road network into a directed weighted graph (DWG), on which the shortest route (route with the lowest weight) is calculated. The weight mining algorithm should combine both static road attributes (such as road level, speed limits, etc.) and dynamic road attributes (such as traffic conditions, etc.) to improve the quality of the planned routes.
For this topic, DiDi provides anonymized road network data to academic community.
For more information, please contact: gaia@didichuxing.com
2. Research on Multi-Sensor-Based Indoor Positioning
2.1 Background
Positioning is the foundation of all LBS services. In DiDi business scenarios, positioning supports many upper-level applications, one of which is passenger pickup point recommendation and walking navigation. A major difficulty of positioning is achieving accurate positioning in indoor areas.
In an indoor scenario, as the mobile device is unable to receive a GPS signal, it is practically impossible to obtain the location via GPS positioning. As such, the location is usually obtained using the network positioning method. Due to the combined low accuracy of base station positioning, the difficulty of estimating the AP location in the room, and the instability of the Wi-Fi signal, it is difficult to achieve high accuracy with simple network positioning.
Nowadays, mobile phones are becoming more popular and powerful. Nine-axis sensors (accelerometers, magnetometers, gyroscopes), barometers and illuminators have become standard equipment for various mainstream mobile phone types on the market, with the accuracy also gradually improving. Therefore, it is possible to optimize indoor positioning effects by using multi-sensors. For instance, users may have a more accurate location provided by GPS before entering the buildings, and users’ indoor location can be obtained by using PDR (Pedestrian Dead-Reckoning) technology to calculate their trajectory after entering buildings.
This project aims to employ multi-sensors to obtain more accurate indoor location, and to carry out various prospective technical exploration in indoor positioning optimization. The goal is to achieve accurate indoor positioning, support the applications of accurate pickup point recommendation, indoor walking navigation, and so on.
Information available, including but not limited to:
1. Accelerometer
2. Magnetometer
3. Gyroscope
4. Barometer
5. Illuminator
6. Wi-Fi fingerprint
7. Cell fingerprint
Other available data, including but not limited to:
1. Map POI data
2. Indoor Maps
2. 2 Research Target
1）Indoor PDR Algorithm
The objective of the indoor PDR algorithm is to use a mobile phone’s nine-axis sensors (accelerometers, gyroscopes, magnetometers) and barometers/illuminometers to estimate the position change of the user and calculate the current position with regard to the previous trusted position. The issues involved in the process include: How to obtain a sizable sample pool and true values? How to determine the user’s initial direction? How to determine the placement of mobile phones? How to estimate the user’s direction and speed of movement? How to ensure the algorithm applies to a variety of phones? How to verify the actual effect?
2） Indoor Fused Positioning Algorithm
The objective of the indoor fused positioning algorithm is to achieve more accurate indoor location by combining the PDR algorithm with Wi-Fi positioning, indoor roads and other information. Issues involved in the process include: How to obtain a sizable sample pool and true values? How to integrate multisource information to optimize positioning? How to verify the actual effect?
2.3 Excepted Outputs
Considering DiDi’s business scenarios, we require that the methods devised via the project possess strong practicality, involving the following aspects:
1) Feasibility—the method must be easy to deploy and apply under existing scenarios and technical conditions.
2) Price-quality ratio—the method must produce enough value within limited investment.
3) Coverage—the method must eventually benefit a sizable number of users.
With the premise that the method does demonstrate strong practicability, it shall be evaluated with regards to the following aspects:
1) Positioning precision—deviation distance from indoor positioning’s true value, with the median calculated by sampling multiple positioning points.
2) Positioning accuracy—deviation distance from indoor positioning’s true value, with the deviation calculated by sampling multiple positioning points.
3) Practicability—the method’s feasibility, price-quality ratio and coverage.
Topic Three：: Voice technology in travel scenarios
1 Health Status Monitoring via Speech
1.1 Background
In-car scenarios are special in that the driver may be subject to negative affections (such as illness and drunkenness), especially for drivers with extended time of travel.
This topic aims to analyze in an immediate manner the health status of the driver via audio and response.
1.2 Research Target
For the above key issues, research objectives include but are not limited to the following:
1） Health Detection
The topic aims to build the ability to detect health-related events via in-travel audio data, with the main focus being on method and performance, to determine the driver’s health via audio data. It may involve procedures including class definition, data set construction, model training and evaluation.
2） Drunkenness Recognition
The topic aims to recognize whether the driver is drunk or not. It may involve technologies including driver-passenger identification, and degree-of-drunkenness recognition.
2. Cross Channel Speaker Verification
2.1 Background
The technology for speaker identification is presently considered mature within the industry. However, the change of speech channels, even from the same speaker, could impact the performance of speaker-ID systems in certain circumstances.
2.2 Research Target
Channel-Invariant Speaker Verification
Algorithms such as the classical i-vector, or later, end-to-end speaker-IDs, are becoming mainstream forms of speaker verification. The topic herein will focus on speaker-ID techniques that are based on cross-channel speech data, which is more complex than single channel speech.
This topic seeks to achieve speaker identification or verification via a cross-channel technique. For instance, the data for enrollment and verification may come from different channels. In other words, this topic is seeking for a channel-invariant speaker verification.
Speech might input from, but not be limited to, the following channels:
1） Close-field high resolution microphone
2） VoIP telephone communication
3） Far-field high resolution microphone
A cross-channel, content-independent speaker verification technique needs to be established with a minimum expected EER.
Topic Four：Computer Vision
1. Research on LargeScale Structure from Motion（SfM）
1.1 Research Area
Computer Vision, Visual Perception, Structure from Motion (SfM), Simultaneous Localization and Mapping (SLAM), Structural Perception and Understanding
1.2 Background
As people’s travel habits evolve, the market has seen a surge in demand for indoor localization & navigation on mobile phones, with indoor localization being essential for related Apps.
Signal strength and accuracy of traditional localization techniques (e.g. GPS) are greatly affected by indoor environments. The manner in which to utilize existing signals for indoor localization is a challenging and practical task. Visual information such as images/videos have the advantages of low data collection cost, no additional equipment requirement, and rich perceptual semantics, and thus have great potential in contributing to localization technology.
1.3 Target
The purpose of this topic is to explore the visionbased indoor localization and navigation schemes and strategies on mobile phones, including: vision–based indoor structure acquisition and construction technology, visual matching technology, and their implementation probability and plan in indoor localization.
Specific implementation involves the following aspects:
1）3D reconstruction data acquisition scheme:
determining the equipment to be used for image data collection, acquisition strategy, and organization of data;
2）Visual reconstruction technology:
recovering the basic 3D model structure of the corresponding scene based on the acquired image;
3）Vision-based localization technology:
locating the position of the camera center in the constructed scene based on one or more images.
1.4 Expected Outputs
Specific research can be carried out according to the above content as the core overall scheme or key content.
The evaluation method of the research content considers the following aspects: cost of data collection; construction accuracy, integrity of model, etc.; accuracy, calculation speed, success rate of visual localization technology, etc.
The field of visual localization technology currently has a sizable amount of relevant research in academia, but its realworld application is still in the exploration stage. This topic is expected to explore and examine the technology’s relevant feasibility and implementation plan. The expected output is as follows:
1）A localization solution based on visual signals. The solution shall describe in detail the means to implement, and the feasibility and potential difficulties of using visual signals for localization, including steps and methods of implementation, strategies of data collection and update. More specifically, it may include: image requirements during model construction, such as number, resolution, shooting angle of images, etc.; the steps associated with model construction, and the requirements for computing resources; computational complexity and storage complexity of visual matching during localization.
2）Core algorithm and prototype system, including: scene structure construction algorithm，imagebased localization model (estimation of current location and direction)，presentation systems, etc.
3）Evaluation indicators of modules, and the evaluation results of the proposed algorithm.
2. End-to-End Learning for Self-Driving
2.1 Research Area
Computer Vision, Robotics
2.2 Background
DiDi Researchers have set a research agenda highlighting three areas that are central to DiDi’s longterm vision and strategy, including experiments in end to end learning for selfdriving cars-i.e. use deep learning to solve perception, behavior prediction, pathplanning and control problems directly from raw sensor input such as Lidar/Camera and GPS/IMU information.
End-to-end learning has become increasingly popular in academic and industry sectors. Particularly within the context of self-driving cars, many recent works have shown promising results in using raw sensor data such as images, lidar inputs and/or radar to generate favorable outcomes.
2.3 Target
We are interested in experimenting with new end-to-end systems for multiple tasks – namely, with only camera images and/or with 3D Lidar inputs, what results can be derived from end-to-end systems for different modules, and what potential limitations exist. For example:
1) Perception: given the image and 3D objects, using a deep network to predict object contours (segmentation), categories (classification), velocity (speed and direction) etc.
2) Behavior prediction: predicting the intention or future of the objects around the autonomous vehicle based on raw sensor data; for example, whether a neighboring vehicle is about to cut in or merge into the current lane, or if a pedestrian is about to cross the crosswalk.
3) Path-planning and control: directly outputting the planned trajectory of the selfdriving car based on input sensor data, i.e. the current and future x, y, positions and the expected speed at each waypoint, or control signals (steering wheel, pedal and brake).
4) Design of the algorithm to be performed in real time, e.g. 10 to 30Hz with given current GPU setup. This involves exploration of deep net trimming and GPU speed up.
2.4 Excepted Outputs
1.A full exploration and evaluation of the capability of end to end learning algorithms on multiple datasets and tasks.
In particular:
1）Experiments on multiple datasets, including public dataset like KITTI, private dataset provided by DiDi, and simulated datasets such as Deep-Drive.
2）Experiments on multiple tasks, e.g. while provided with raw sensor data, using the deep learning network to perform tasks such as localization, object detection, semantic level segmentation, instance level segmentation, object velocity/behavior prediction, and robot path planning including trajectory and final control signal output. Analyze the strengths and weaknesses of the end-to-end algorithm on each stage, and successful/failed examples.
3）Experiments on different sensor inputs, e.g. a combination of LiDar, Radar, Camera inputs, with or without GPS/IMU input.
2.Analysis of these algorithms in technical report format.
3.Related source code to reproduce the analysis and experiments.
4.With some imposed limitations and after removing certain sensitive data and algorithms, it is possible to aim for publications on top conferences such as CVPR/ICCV or journals like PAMI.
3. Multi-Person Abnormal Behavior Detection
3.1 Research Area
Computer Vision, Instance Segmentation, Human Key-Point Detection (Pose Estimation, Human Skeletal System Key-Point Detection), Action Recognition
3.2 Background
As techniques in machine learning and computer vision are developing rapidly, a wide crosscity network of cameras has been set up to construct a powerful surveillance system which guarantees social security. Using this camera network in conjunction with advanced image analysis techniques, a powerful system can be built to recognize passengers and their actions, thus to discover abnormal human behavior in certain environments such as elevators or vehicles, and warning alarms might be triggered subsequently. This can reduce accidents and protect the safety of people’s lives and property.
3.3 Target
1）Instance Segmentation
The prediction of future events is an important feature of intelligent behavior, and image prediction is one task falling under such goals. Recent work shows that the semantic prediction of future frames for semantic segmentation is more effective than performing prediction of RGB frames and segmentation separately. Therefore, it is promising to explore a new solution for instance segmentation to alleviate conflicting predictions in a cabin environment.
2）Human Skeletal System KeyPoint Detection
Human skeletal system keypoint detection plays an importance role in describing human pose and predicting human behavior, which is the foundation of many computer vision tasks such as action recognition, abnormal behavior detection, and autonomous driving. Through the detection and analysis of key-points, abnormal behavior recognition and prediction can be leveraged to guarantee the safety of passengers.
3）Predicting Conflicts
Predicting conflicts through machine learning is becoming a popular topic in the intersecting fields of sociology and computer sciences. Due to the uncertainty and variability of factors that lead to conflict, conflict prediction is still very controversial within the academic field. With the increase of data resources and computational power, image analysis techniques can be used to predict the probability of conflict between drivers and passengers.
4. Image Enhancement
4.1 Research Area
Computer Vision, Image Enhancement, Image Deblurring
4.2 Background
Images and videos are playing an increasingly important role in the field of public safety. More and more surveillance cameras are installed in cities to ensure public safety. These cameras are installed in corporate environments, living quarters, markets, and individual locales etc. Due to the myriad of affecting factors in the shooting environment and the imaging quality of the imaging device, the generated images and videos are often blurry and unclear, which imposes great difficulty and challenge for subsequent image recognition. Blurring can be caused by many factors, such as insufficient night light, low resolution of imaging equipment, extreme weather conditions (rain and fog, etc.), excessive exposure of images/video, inaccurate camera focus, rapid movement of objects, etc.
4.3 Target
To solve the above key issues, Our research objectives include but are not limited to the following:
1）Dark Image Enhancement
Dark image enhancement mainly resolves scenes where the image is filmed at night or in a dark area. We plan to design an end-to-end model based on deep neural networks to generate clear images through dark image enhancement technologies, and ultimately achieve state-of-the-art performance on open image datasets.The technologies will satisfy all requirements in practical test datasets.
2）Image Blurring
Image blurring has been a persistent issue in the field of image processing. The reasons for image blurring run a wide gamut of complex factors, such as camera shake, imprecise focus, high speed movement of objects, and so on. We plan to design an endtoend model based on deep neural networks to generate clear images through image enhancement on blurred images, and ultimately achieve state-of-the-art performance on open image datasets.In the practical test datasets. The technologies will satisfy all requirements in practical test datasets.
Topic Five：Machine learning
1. Supply and Demand Forecast
1.1 Research Area
Spatiotemporal Data, Crowd Transfer Analysis, City Trip
1.2 Background
The main purpose of the city trip platform is to connect drivers and passengers. The means by which to effectively use data, through big data and machine learning technology, to display and model supply and demand in order to achieve supply-demand balance under high-frequency travel is the key to improve the efficiency and service of the entire platform.
By establishing a unified real-time forecasting system, we seek to support a variety of business scenarios such as order dispatching, driver scheduling, diversion, pricing, and product strategy, and improve the efficiency of the entire trip platform.
1.3 Target
1.3.1 Spatiotemporal Data Forecasting
This topic aims to establish the spatio temporal data prediction capabilities on the city trip platform, which involves dataset organization, feature engineering, model training, and verification. The prediction of spatio temporal data is the main challenges of the analysis of spatiotemporal. We intend to establish an evaluation system and have some requirements as follows:
1) System needs to provide a real-time prediction
2) System should adapt to some unexpected situations, such as holidays, extreme weather, unexpected events, etc.
1.3.2 Crowd Transfer Analysis
The fundamental function of the city trip platform is crowd transport. Only by deepening the understanding of crowd and behavioral characteristics can we take a more scientifically advanced approach to fulfill the platform’s goals.
2. On-Demand Ride-Sharing Algorithm
2.1 Research Area
Online Trip-Vehicle Assignment; Improve Efficiency of Ride-Sharing System
2.2 Background
The rapid development and popularization of online ride-hailing services has largely improved the efficiency and quality of on-demand mobility. Traffic congestion has been relieved through vehicle sharing. Real-time dispatching and optimization is one of the key points of a big-data-enabled smart city. Ride sharing services improve the system’s efficiency by combining multiple trips into one. These services aim to be efficient, cheap and environmentally friendly.
Combining more orders with less experience loss will also greatly improve carpooling efficiency.The realization of carpooling efficiency depends on pricing power, demand structure and ordering ability, with all parts being mutually interdependent. The pricing strategy can inform basic pricing policies and subsidy allocation schemes based on different objectives. Meanwhile, the underlying structure of mobility demand can be significantly heterogeneous across time and space. Therefore, given the differences in underlying structure, the volume of demand can be significantly different when aiming to achieve similar efficiency. When provided with demand, the dispatching module performs real-time trip vehicle assignment decisions and improves systematic efficiency through evolution of optimization algorithms.
The comprehensive effects of these three components decides the efficiency of a ride-sharing system. Quantitative description of their independent influence and interactive relationship can effectively improve the overall framework performance.
2.3 Target
2.3.1 Optimization of Carpool Order-Dispatch Decision-Making in Time-Series
One major characteristic that differentiates online carpool services from other transportation models is its quick response to real-time mobility demands, which means tremendous demandmust be fulfilled in a very short time. Therefore, the key of order dispatching decision is striking the complicated balance between the best solution in the current matching round, and the possibility of a better solution in the future. In other words, a tradeoff between a short term, “greedy” solution or a solution which can wait to reap potential delayed benefits.
2.3.2 Modeling of the Coupling Between Demand Structure and Order Dispatch
Within the current order-dispatch scheme, there exists different elements like basic order–dispatch algorithms, optimization of decisionmaking in time series, control of attributes related to customer experiences, etc.
With a different underlying demand structure, these factors’ interactive relationship with ridesharing efficiency will change accordingly. Therefore, a flexible dispatching framework which is adaptive to different demand structures can be of great significance.
2.3.3 Demand Structure Modeling
The underlying demand structure can be heterogeneous especially when considered with regard to different objectives. Effective structure adjustments can be important for operations goals, which makes research topics, such as route mining generalization and route correlation, of particular interest.
2.3 Detour Recognition
Real-time identification of carpool detours includes identifying whole order carpooling detours, suborder detours, etc. Generally, this process involves the definition of abnormal problems, user feedback collection and management, feature data set construction, and model training and verification. Among them, dismantling and modeling of the problem is the most difficult. There are numerous connotations in a carpool bypass scenario, and the following three requirements should be satisfied in modeling machine learning.
1）Cohesiveness
Each sub problem scene has a clear definition, and the boundary between any two subscenes is as clear as possible. The resolution and definition of each subscene is highly cohesive.
2）Overall situation
The correlation between the sub problems and the relationship between the subscene and the overall scene will have a strong influence on the discovery of the problem and the timely resolution of early problems. It needs to be considered as a whole.
3）Modeling indexes
The model needs to consider the accuracy, recall rate and the balance between cost and user experience, and give consideration to the interpretability of the model.
2.4 Matching Rule
The goal is to build a carpool scheduling strategy based on deep learning and deep reinforcement learning, which takes into account the characteristics of order time, cost, mileage, road condition, spare seat, driver attribute, passenger attribute and so on, and combined with the experience of cost and benefit, thereby enhancing resource utilization and driver passenger experience.
1）Multi-objective optimization
We should fully consider the interests and constraints of drivers, passengers and platforms, so as to achieve the optimization of the ride experience, the balance between experience and efficiency, and long-term growth of the platform.
2）Modeling indexes
On the basis of fully understanding the business and data, we should build a generally flexible and configurable scheduling model to drive for a marked increase in core business indicators, such as a lower detour rate.
3. Order Dispatch in Large-Scale Online Systems
3.1 Research Area
Combinatorial Optimization, Reinforcement Learning, Multi-Objective Optimization, Online System
3.2 Background
In recent years, we have witnessed a rapid growth on the emerging on-demand ride-hailing services. Compared to the traditional cruising taxi services, on-demand ride-hailing can ensure better efficiency and user experience from the perspective of a platform. Order dispatch is one of the key algorithmic components in ride-hailing platforms.
In real-world scenarios, an order dispatch algorithm should meet the following requirements:
1）Run in an online, real-time fashion.
Take the Drip platform as an example, the daily order volume of the platform is 30 million, which poses high requirements for the real-time responsiveness and complexity of the algorithm.
2）Balance the trade off between platform efficiency and user experience, etc.
Due to the complexity of the task, online order dispatch is impossible through human power alone; it is highly dependent on algorithm designing. From another perspective, the dispatch algorithm is also good experimental soil for research fields such as artificial intelligence, machine learning, and operation optimization. In summary, the issue of dispatching can be one of the more important applications of the converging academic and industrial circles.
3.3 Target
3.3.1 Order Dispatch Framework
An important feature of the order allocation problem is its network effect and timing effect.Specifically, vehicle network is a bilateral market, and the matching between the divisions is not independent of each other. It forms a network structure in time and space. At the same time, it is necessary to take into consideration various factors such as platform efficiency, driver experience, fairness, etc. while dispatching, and form a multi-objective optimization task. Therefore, it is of great value to design a matching framework that satisfies the above conditions, in order to balance optimization goals and constraints, and integrate time series and space matching requirements.
3.3.2 Personalize Order Dispatch
Perform order dispatches according to the user’s real-time status and user characteristics to improve efficiency.
3.3.3 Fairness and Interpretability
DiDi’s online ordering system serves a large number of drivers and passengers every day. Target customers are individuals with emotional appeals. Therefore, the interpretability and fairness of the algorithm are often as important as the platform efficiency and experience. Introducing fairness into optimization goals or constraints and generating matching criteria that can be interpreted and backtracked must be taken into consideration.
4. Design of Evaluation System for Online Car Hailing
4.1 Research Fields
Deep Learning, Machine Learning, Semi-Supervised Learning, Active Learning, Labeling Approach, Transportation, User Experience
4.2 Background
With the disruptive changes in the mobile Internet, the online carhailing industry has risen globally. Didi, the world's leading online car-hailing and smart transportation platform, has been profoundly changing people's travel and even lifestyle.
In order to increase income for drivers that produce quality service, designing an evaluation system for ride sharing services has also become key to the construction of the driver-passenger ecology, taking into account the passenger experience, the fairness of the driver's order and the reasonable income level of the platform. DiDi's current evaluation system is based on machine learning modeling. Accurate labeling data of feedback from drivers and passengers and continuous optimization of the model will also enable the design of the network car evaluation system to be more perfect, thus further enhancing the user experience of drivers and passengers alike.
4.3 Target
4.3.1 Labeling Approach Research
We aim to build an automatic or semi automatic sample labeling system to provide sufficient samples for model training. It requires a combination of manual labeling of samples related to semisupervised learning, active learning, etc. How to ensure that the quality of auto-labeled samples is as good as those of manually labeled is the most difficult problem. In doing so, the following two requirements must be met：
1) Accuracy
Each sample must be evaluated accurately to achieve or approximate the quality of manual labeling.
2) Automatic or Semi automatic
Sample production must be automated or semi automated so that it is able to meet the requirements for large scale model training like DNN.
4.3.2 Model Algorithm Research
We explore appropriate methods including machine learning, deep learning, and enhanced learning to continuously improve on model performances.
Specific indicators for the evaluation are as follows:
1）AUC
2）Precision
3）Recall
4）Engineering Architecture
5. The Study of Fault Determination for Cancellation in Ride-Sharing Business
5.1 Research Fields
Deep Learning, Machine Learning, Semi-Supervised Learning, Active Learning, Labeling Approach, Intelligent Transportation, User Experience
5.2 Background
With the expansion of the mobile internet, ride sharing business has thrived across the globe. A good ecology of drivers and passengers is of utmost importance. Take an ordinary ride scenario as an example. When the passenger enters his own destination and clicks “call ride”, the platform will match the appropriate driver in real time. After the matching is complete and before the billing starts. the driver or passenger can cancel the order at any time, but they may face certain predication on responsibility, since an act of cancellation from one party will often affect the experience of the other party.
This impartial judgment call is currently achieved by machine learning modeling. When a driver or passenger cancels an order from either party, the system can determine the cancellation responsibility of both parties in real time and automatically give immediate feedback to effectively protect the rights of both parties. The accurate labeling of the feedback sample from drivers and passengers and the improvement of the machine learning model algorithm can make the responsibility judgment of the cancellation behavior more accurate, thus further improving the user experience of drivers and the passengers.
5.3 Target
5.3.1 Data Labeling
To formulate the ability to label data automatically or semi-automatically and provide enough samples for model training. Generally, the methodology includes manual labeling, semi-supervised learning and active learning. In general, it is difficult to achieve the same standard of quality in automatic labeled data as data labeled by humans. The following two standards need to be assured:
1）Precision
Automatic labeled data should be the same quality as data labeled by humans.
2）Automation
Data needs to be labeled automatically or semiautomatically to maintain a substantial volume (billions in units).
5.3.2 Model Algorithms
Exploring additional model algorithms like machine learning, deep learning and reinforcement learning to achieve better precision/recall performance. More specifically, evaluation metrics include:
1）AUC
2）Precision
3）Recall
4）Engineering Complexity
6. Research on Prediction Model for Individual Transportation Needs
6.1 Research Fields
Prediction on Individual Transportation Needs
6.2 Background
DiDi is an on-demand transportation service provider, aiming at fulfilling individual transportation needs. After a passenger types in their destination on the DiDi mobile app, the task for DiDi is to match the featured needs with a nearby transportation service provider precisely and efficiently.
It is of great significance to study and predict the transportation needs of DiDi users within a certain timeframe so that the operating team can manage the transportation sources and guide the drivers accordingly to improve the transportation sharing service. For example, by developing a workable algorithm predicting the probability that a user may require transportation service from A to B within 24 hours, DiDi can provide the user with personalized recommendations before their trip, with emphasis on the features of said upcoming trip (i.e., likelihood of said trip being made, type of destination). User experience can be improved with efficient operating strategies.
Usually, the decision-making process when a certain user chooses a transportation service is partially revealed on DiDi platforms. User history data is highly sparse in nature, although it is comprehensive to make predictions on future transportation demands on a personal level, since the probability of future trip of A to B is affected by numerous features such as weather, global demand, etc., the users’ transportation behaviors do have inherent patterns.For example, a shuttling service between home and workplace typically exhibits low variation in time, while a typical travelling transportation is more likely related to a landmark. Given enough user behavioral data in transportation, their future demand within a certain period is predictable.
Developing and implementing machine learning and data mining algorithms to predict the future trips of individual users and creating correspondingly operational strategies are essential for DiDi to achieve its business objectives.
6.3 Target
Predicting the Upcoming Individual Trips within 24 Hours
To predict future routes an individual user may take, using service from DiDi, within 24 hours, and to calculate probabilities, through building machine learning models based on the big data in transportation from DiDi, including history trips, location, and orders, and incorporating POI, weather, and events (Concerts, Holidays, Chinese New Year).
6.4 Expected Outputs
Model Evaluation:
1) TOPN the precision/recall of true routes in predicted routes (typically, N=1,3)
2) AUC of predicting model for TOP1 route. It may vary based on the nature of the trips. E.g. regular shuttling trips between home and workplace are easier to predict compared with trips of other types.
Requirements:
1) For a certain type of trips, the precision/recall of the prediction model on individuals'future trips should be sufficiently high as to ensure practical application.
2) The model is an advanced innovation.
3) The model/algorithm is considered practical for DiDi's business operations.