FADACS: A Few-shot Adversarial Domain Adaptation Architecture for Context-Aware Parking Availability Sensing
Wei Shao, Sichen Zhao, Zhen Zhang, Shiyu Wang, Mohammad Saiedur Rahaman, Andy Song, Flora Dilys Salim
FFADACS: A Few-shot Adversarial Domain AdaptationArchitecture for Context-Aware Parking Availability Sensing
Wei Shao ∗ Sichen Zhao ∗ [email protected]@rmit.edu.auSchool of Science, RMIT University Zhen Zhang
School of Science, RMIT [email protected]
Shiyu Wang
School of Science, RMIT [email protected]
Mohammad Saiedur Rahaman
School of Science, RMIT [email protected]
Andy Song
School of Science, RMIT [email protected]
Flora D. Salim
School of Science, RMIT [email protected]
ABSTRACT
The existing research on parking availability sensing mainly relieson extensive contextual and historical information. In practice, itis challenging to have such information available as it requirescontinuous collection of sensory signals. In this paper, we designan end-to-end transfer learning framework for parking availabilitysensing to predict the parking occupancy in areas where the parkingdata is insufficient to feed into data-hungry models. This frame-work overcomes two main challenges: 1) many real-world casescannot provide enough data for most existing data-driven models.2) it is difficult to merge sensor data and heterogeneous contextualinformation due to the differing urban fabric and spatial character-istics. Our work adopts a widely-used concept called adversarialdomain adaptation to predict the parking occupancy in an areawithout abundant sensor data by leveraging data from other areaswith similar features. In this paper, we utilise more than 35 millionparking data records from sensors placed in two different cities, oneis a city centre, and another one is a coastal tourist town. We alsoutilise heterogeneous spatio-temporal contextual information fromexternal resources including weather and point of interests. Wequantify the strength of our proposed framework in different casesand compare it to the existing data-driven approaches. The resultsshow that the proposed framework outperforms existing methodsand also provide a few valuable insights for parking availabilityprediction.
CCS CONCEPTS • Computer systems organization → Embedded systems ; Re-dundancy ; Robotics; •
Networks → Network reliability. ∗ Both authors contributed equally to this research.Permission to make digital or hard copies of all or part of this work for personal orclassroom use is granted without fee provided that copies are not made or distributedfor profit or commercial advantage and that copies bear this notice and the full citationon the first page. Copyrights for components of this work owned by others than ACMmust be honored. Abstracting with credit is permitted. To copy otherwise, or republish,to post on servers or to redistribute to lists, requires prior specific permission and/or afee. Request permissions from [email protected].
Woodstock ’18, June 03–05, 2018, Woodstock, NY © 2018 Association for Computing Machinery.ACM ISBN 978-1-4503-XXXX-X/18/06...$15.00https://doi.org/10.1145/1122445.1122456
KEYWORDS transfer learning, parking availability, sensor networks
ACM Reference Format:
Wei Shao, Sichen Zhao, Zhen Zhang, Shiyu Wang, Mohammad SaiedurRahaman, Andy Song, and Flora D. Salim. 2018. FADACS: A Few-shotAdversarial Domain Adaptation Architecture for Context-Aware ParkingAvailability Sensing. In
Woodstock ’18: ACM Symposium on Neural GazeDetection, June 03–05, 2018, Woodstock, NY.
ACM, New York, NY, USA,10 pages. https://doi.org/10.1145/1122445.1122456
Parking availability sensing plays a vital role in urban planningand city management [24, 37]. According to a recent study, driversspend more than 100,000 hours per year in looking for parkingtheir cars [27]. Moreover, seeking for available parking can leadto severe traffic congestion and air pollution [9]. Hence, effectiveparking availability sensing can help drivers find a vacant parkingspot. This also helps government to take appropriate measure byunderstanding the utilisation of parking facilities and provide moreon-street parking lot in the areas with high parking demand.Parking dynamics have been studied in many research domains.In recent times, two types of sensing systems (i.e. explicit and im-plicit) have been used to infer parking availability around the cities.The explicit sensing systems take a direct approach to measure theparking occupancy through physical sensors such as undergroundsensors, RFID sensors, and monitoring cameras. In contrast, im-plicit sensing systems use an indirect approach to measure parkingoccupancy, e.g., through the sensing of contextual information suchas weather, the number of restaurants and office building nearby,and density of public transportation stops. [34].Most existing data-driven solutions heavily rely on the long-termand historical data which is not always available in the real-worldscenarios [23]. In recent times, several works leverage the transferlearning techniques to estimate the traffic in areas without muchhistorical data [31, 32]. However, domain shift and unsupervisedlearning remain as two main challenges in these parameter transfer-ring models. Another common challenge is that most of the existingworks focus on the temporal dependency of the contextual informa-tion and parking records. However, spatial dependency also playsa key role in parking occupancy because the status of a parkingslot is highly correlated with nearby parking slots. For example, a r X i v : . [ c s . L G ] J u l oodstock ’18, June 03–05, 2018, Woodstock, NY anonymous author, et al. drivers tend to park in a spacious space rather than a narrow spacesince crowed parking spaces is likely to rise the parking difficulty,and they also prefer a low occupancy area because of the Nashequilibrium [10]. Therefore, considering both spatial and temporaldependency is essential to parking occupancy prediction problem.Adding to the above challenges is the highly diverse featurespace in the source and the target domain when the sensor dataare collected from two really different cities. This paper, in particu-lar, presents a challenge that is often present when sensors in thedifferent cities are deployed by local authorities and the data arecollected by different agencies, capturing local contextual informa-tion that is pertinent to the local urban fabrics with their specificcharacteristics. In this paper, the source city is the city centre of anAustralian state’s capital with its Central Business Districts (CBD)areas, and the target city is a little coastal town mainly populatedby retirees and is very popular with tourists, in particular whenthe weather is clear. Hence, the parking patterns across the tworegions are highly diverse and may not be directly transferable.To overcome the challenges of integration of spatial dependencyand temporal dependency and shared features extraction, we de-sign a domain adaptation architecture called FADACS which canlearn the parking occupancy without much historical parking databy utilising contextual sensor and parking sensor data from otherareas. We use the idea from computer vision area [30] and incor-porate with meta sensing technologies. Specifically, we proposea generative adversarial networks-convolutional long short termmemory model for parking occupancy prediction by combininggenerating ability of generative adversarial networks (GAN) andspatio-temporal forecasting ability of convolutional long short termmemory (ConvLSTM). Compared to existing transfer learning mod-els such as parameter transferring models [31], GAN-based transferlearning work can easily learn the shared features of source domain(Where historical data is available and rich) and target domain(where we would like to predict the parking occupancy with nohistorical parking data) using adversarial learning mechanism, andit does not need historical data from target area. Additionally, Con-vLSTM model applies the convolution operations on the spatialdomain and recurrent layers to the temporal domain [31]. We em-bed such model into our adversarial learning framework and test iton two different real-world parking dataset with contextual infor-mation. The experimental results show that our proposed modeloutperforms other existing transfer learning models. We also showthat the contextual information has a significant influence on theprediction accuracy. In particular, the contributions of this paperare as follows: • To our best knowledge, We are the first to propose a GAN-based spatio-temporal transfer learning framework to pre-dict the parking occupancy in areas without historical park-ing records by utilising parking data from other areas andcontextual information. We compare our proposed modelwith traditional transfer learning model which only taketemporal information into consideration and state-of-the-artworks such as ConvLSTM which consider both spatial andtemporal information but only use parameter transferring ap-proach to learn the distribution from the source domain. Theexperiments validate that our work which incorporates both spatial information and temporal information and leveragesthe GAN-based transfer learning framework can improvethe parking prediction accuracy. • We conduct an in-depth analysis of contextual factors whichhave potential influences on parking occupancy in differentregions. We conduct the quantitative investigation on park-ing sensing by both implicit and explicit ways. Our studyfound insights on the contextual factors that have potentialinfluence on parking occupancy.The rest of the paper is organised as follows. Section 2 describesrelevant works in parking sensing area and transfer learning area.Section 3 introduces parking and contextual information dataset.Section 4 shows our data preprocessing pipeline. Section 5 illustrateour proposed framework. Section 6 provides experimental results,followed by conclusion in Section 7.
Parking prediction:
Parking availability predictions, which canbe treated as one of the time-series issues, are appropriate for manymethods. Yu et al. [36] verify the effectiveness of making real-timeparking availability prediction using time series model. They estab-lished a variant of the autoregressive integrated moving average(ARIMA) model to predict remaining berths in an undergroundparking lot at Xinjiekou in Nanjing, China. Pfl (cid:220) ugler et al. [22] makea detailed analysis of the importance of publicly available infor-mation. The authors train a neural network (NN) model based oncontextual features solely, and it shows that the prediction madewithout the historical parking data could be very effective and thethree most common information: time, location and weather havethe greatest impact compare to all other features. Except the linearmethods, Chu et al. [21] adopt the backpropagation neural network(BPNN) model proposed by Haviluddin et al. [8] on available park-ing spaces data collected in Xi’an, China. BPNN makes a nonlinearmapping between inputs and outputs, and the results show that itcan generate effective predictions for parking lots with differentcapacities. Shao et al. [25] further utilise a large real-world parkingspaces dataset in Melbourne, Australia and train a long short-termmemory (LSTM) model on that dataset. Results are quite promising.
Domain transfer learning:
Since most of the machine learningmodels assume that the overall of training and test data are IID(independent and identical distributed), which is not always thecase in the real world [20]. One major drawback causing by thisissue is that the test data which comes from a shifted distributionmostly will lead to an unexpected performance drop. Except forthe traditional approach, which is to build a new model and re-train that model, transfer learning is widely used to overcomethis problem due to its better performance and efficiency. Transferlearning enables us to learn knowledge from the source domainupfront and apply that knowledge to a new, relative data or targetdomain. Transfer learning has also developed different areas suchas lifelong learning [28] and multitask learning [2] and transductivetransfer learning [1].Pan and Yang [20] considered transductive transfer learning issimilar to domain adaptation, which has the ability to transfer theknowledge from the labelled source domain to unlabelled targetdomain. Generative Adversarial Networks (GANs) [6] since 2014
ADACS: A Few-shot Adversarial Domain Adaptation Architecture for Context-Aware Parking Availability Sensing Woodstock ’18, June 03–05, 2018, Woodstock, NY has become one of the hottest concepts in the field of artificialintelligence. Ganin et al. [4] first added adversarial mechanism intodomain adaptation and propose a new framework called DANN(Domain-Adversarial Neural Network). The objective of this methodis to generate features which contribute most to classification whilemaking the discriminator unable to determine the source of thosesamples. In addition to mapping source and target domain into thesame feature space, fine-tuning is another area of transfer learning,which is a process to reuse the training model into a second similartask and also is a simple method to transfer knowledge. Yosinski et al. [35] first demonstrated the transferability of features froma neural network. Since then, fine-tuning has been widely usedin multiple areas. Google BERT [3] is considered as a milestonein NLP, which comprises a pre-training stage and a fine-tuningstage. Similarly, to detect pathological brain in magnetic resonanceimages (MRI), the authors of [11] achieved higher performance byusing parameter-transfer learning based on pre-trained AlexNetmodel. Besides the improved performance and the reduced trainingtime, another advantage of the fine-tuning is that it can counterthe over-fitting problem, which usually occurs on small datasets.Facing ’cold-start’ problem on expending market into a new city,Guo et al. [7] gives a new framework, called CityTransfer, whichcould learn the knowledge of inter-city and intra-city, based oncollaborative filtering. One of the latest work applied few shotlearning technique in sensing area is proposed by Gong et al. [5].They learn the behaviours of each user only with a few samplesusing transfer learning technique. However, they did not apply themodel to spatio-temporal data and contextual information.In this paper, due to the lack of historical data on the targetdomain, we choose to approach the problem with domain adapta-tion, allowing transfer of knowledge from the source domain to thetarget domain.
The two cities that are being investigated in this research are theCity of Melbourne and the town of Rye. Both are in the state ofVictoria, Australia. Melbourne is the capital city of Victoria. Themunicipality of Melbourne, with an estimated of 178,955 residents[12], has nearly 1 million people on average per day, visiting themunicipality for work, education, and travel or tourism. On theother hand, Rye is a little coastal town, part of the MorningtonPeninsula Shire municipality. Rye has a population of approximately8,416 in the 2016 census and is located about 100km from the City ofMelbourne. The Mornington Peninsula Shire hosts about 7.5 millionvisitors per year [26], and about 50% of those would visit Rye asone of their destinations, requiring parking spot, as driving is themain mode of transport to get into these coastal areas. Therefore,the major datasets in this paper include parking sensor data, Pointsof Interests (POI) data, weather data, and geographical data.All the datasets used in this system come from the followingplatforms: the City of Melbourne Open Data [13], Time and DateAS [29], Google Map API, and a proprietary Mornington PeninsulaShire data platform.
Parking relative data is from the City of Melbourne Open Data [13].We use the following data sets: • On-street Car Parking Sensor Data 2017 [17] • On-street Parking Bays [19] • On-street Parking Bay Sensors [18]In this section, we will use the name Parking Sensor Data, Poly-gon Data and Location Data to represent each of the above datasets.
The Parking Sensor Data has 35.9 mil-lion records of 2017 on-street car parking in Melbourne Vic, con-taining 35 areas, 5044 sensor devices and 4695 parking slots. Thereason of inconsistency of the sensor devices and parking slots isthat if a sensor device needs to be removed for faulty, low batteryand upgrade the firmware, then a new sensor device with differentid will be replaced. In addition, Open data mentioned that theseare streaming data. Namely, no matter whether the parking slotis occupied or not, each sensor will run the whole day and is con-tinuously generated records. If the parking slot is occupied, thecorresponding sensor will record the arrival time and departuretime. Otherwise, during some periods, the sensor will automaticallyrecord the times and label it as non-occupied. Additionally, everymidnight all sensors will do the record and restart recording again.The detailed format of Parking Sensor Data is shown in Table 1.
There are 24074 records in this dataset, includ-ing all parking slots in the Melbourne area. Each record contains aseries of locations that define the actual boundary of a parking slot.Although each polygon has its unique parking bay Id, only a smallportion of them have a sensor built-in with a street marker Id thatcould link to the parking data mentioned above.
Since the polygon data contains the boundaryof all parking slots no matter whether they have its dedicated sensoror not. In this paper, we also used another data called Location Datathat contains a single longitude-latitude tuple for each parking slotwhich could eliminate ambiguous distance calculation.According to the recommendation from the City of MelbourneOpen Data Platform [13], Location data and Polygon data shouldbe joined by Street Marker Id.
This data is collected by the Mornington Peninsula Shire whichincludes 179288 records cross 527 devices or parking slots in Rye,Victoria. The time range of this data is from 17th Nov 2019 to 20thFeb 2020 and spatially spread in 7 sectors. Details can also be foundin Table 1, and an example of the status for those parking slots isshown in Fig 1.
The Point of Interest (POI) data of Melbourne coming from theCity of Melbourne’s Open Data Platform [13] under project CLUE(Census of Land Use and Employment). It records comprehensiveinformation about land use and updated frequently. We choosethree sub-datasets that covers most of the possible POI categoriesthat related to parking prediction: • Bars and pubs, with patron capacity [14] oodstock ’18, June 03–05, 2018, Woodstock, NY anonymous author, et al.
Table 1: Comparison of the format for both Parking Data used in this paper
Column Dataset DescriptionMelbourne RyeDevice Id √ √
The unique id for the parking sensors.Arrival Time √ √
Date and time that sensor detected a vehiclelocated over it.Departure Time √ √
Date and time that sensor detected a vehicle nolonger located over it.Duration √ √
Time difference between arrival time anddeparture time events, measured in seconds.Overstay Duration × √
Time that a vehicle overstay, measured in secondsIn Violation √ ×
Boolean value, indicate whether the parking eventis violation or not.Street Name √ √
Name for the street of sector that a sensor locatesStreet Id √ ×
An unique Id for streets.Street Marker √ ×
An unique Id for each parking slotDevice Name × √
Name for each device.Sign/Restriction √ √
Parking rule/sign in effect at the time of theparking event.Longitude √ (via Location Data) √ The longitude of the parking sensor.Latitude √ (via Location Data) √ The latitude of the parking sensor.
Figure 1: The location and status example of parking slots located in Rye. The green ones indicate available slots while thegrey ones stand for the slots that currently occupied. • Cafes and restaurants, with seating capacity [15] • Landmarks and places of interest, including schools, theatres,health services, sports facilities, places of worship, galleriesand museums [16]
The first two datasets record allbusiness establishments for pubs, bars, cafes and restaurants. Thedata collection of this part starts in 2002 and updated annually. Wecombine them due to their similar structure, after filtering out thedata in 2017, we get 263 and 3563 records, respectively. Each recordcontains the trading name for that business establishment, an streetaddress and a coordinate which can be pinned point on the map.
The structure of the lastdataset, landmarks and places of interest, is different from the othertwo, which only have 242 records with coordinate information andtheme information. There are 49 themes such as hostel, cinema,library and casino.
Weather data of two places are both collected from Time and DateAS [29]. We gathered the weather data for both Melbourne andMornington according to the time range of the data we gatheredfrom those two places, respectively. Detailed columns used in thispaper are shown below: • Time: The specific time with the weather information
ADACS: A Few-shot Adversarial Domain Adaptation Architecture for Context-Aware Parking Availability Sensing Woodstock ’18, June 03–05, 2018, Woodstock, NY • Temp: The temperature in Celsius scale • Weather: There are 29 different types of conditions, for ex-ample, ’Broken clouds’, ’Clear’, etc. Data in this column isthe combination of weather conditions. • Wind: The wind speed measured in km per hour • Barometer: The barometer in Millibar Pressure Unit
The first important step in pre-processing is to match all parkingslots with the correct location coordinate. We decide to use theStreet Marker Id as the primary key for matching since the DeviceId of a parking slot could be changed. For parking slots in theMelbourne area, we join the three aforementioned files on theStreetMarker column. If the location coordinate is missing for aparking slot, it will be filled with the centre of its polygon. To ensurethe validity of this method, we have proved that all existing locationvalues fall into their corresponding polygon boundary. For parkingslots in the Mornington Peninsula, there is no need for furtheroperation since all slots have a corresponding location coordinate.
In this paper, instead of the individual parking slot, we decide touse the parking lot as the base unit in our experiments. The parkinglot is a cluster of parking slots which fall into the same locality andshare the same parking restriction rule. The size of a parking lot isrelatively small due to its definition mentioned above. By groupingthem together, we switch the objective of the model from predictingwhether a given parking slot is occupied or not at a certain timestepto the occupancy rate of a group of parking slots. Since the statusof a single parking slot can be noisy, this approach simplifies theproblem while still maintain the original goal of parking availabilitysensing.
Since each parking lot are considered asthe base unit in later stages, we first need to ensure that a consis-tent parking rule is shared within this lot. Due to some reasonssuch as construction, the rule could be changed during this period.The parking rules for October, November and December 2017 arereplaced with the whole-year rules. We then create an initial group-ing which groups those spatially connected slots according to theirpolygon boundaries. However, this initial grouping result whichuses the geometric information solely still needs improvements. Asshown in Figure 2a, some of the parking slots are not under thesame parking lot, although they are close to each other and havingthe same parking restriction. Based on this finding, we perform thegrouping operation based on three criteria: connection, distanceand rules. Namely, if the spatially connected, they will be clusteredinto the same lot. For those that are not connected, if they have thesame parking restriction and the distance between them is under athreshold, they will also be put into the same lot. To calculate thisthreshold, we select a specific area and set the value as the sum ofmean connection distance and 1.5 times its standard deviation. Theexample grouping result is shown in Figure 2b, and we cluster all4192 parking slots located in the Melbourne data into 912 separateparking lots.
It is much simpler to cluster parking slots for Ryedataset. Although there is no polygon information in that dataset,each parking slot in the Rye Parking dataset has a coordinate withconsistent parking restriction information. We first group the databy the sector and rule information to reduce complexity. For eachgroup in the same sector with the same rule, we check the distancefor the distance between two neighbour groups, if the distance issmaller than the threshold that we use in the Melbourne dataset,we combine them and get a larger group.
In order to extract the occupancy of a parking lot at a given time,e remove the records that have no vehicle presented or belong toone of the following anomalies: • DurationSeconds is non-positive which is usually caused bya faulty sensor; • ArrivalTime and DepartureTime are both at midnight ex-actly; • DepartureTime is past the midnight of the ArrivalTime; • The records overlapping with other records which could becaused by other unexpected interference.We eliminate over half of the Parking data in this cleaning process.Then, we slice the data every 1 and 5 minutes and calculate theoccupancy of each parking lot at that time.
Based on the POI and Weather dataset that we collect, we calculatethe a series of contextual features which is shown in Table 2.Since the weather condition is a categorical feature, and nor-mal weather may not has a significant impact, we create a binaryindicator for the presence of extreme weather condition (’Dust-storm’, ’Extremely hot’, ’Fog’, ’Hail’, ’Haze’, ’Heavy rain’ or ’Lots’).Then, we match each sample with the most recent weather recordto get the temperature, humidity, extreme weather and barometerfeatures.For the Point Of Interest features, we consider that the totalnumber of POIs and the number of opening POIs within a givendistance may have a higher impact on the occupancy of parkinglots. Since if there are many restaurants around a parking lot, thislot should be more popular during mealtime and have a loweroccupancy rate at other periods.There are a total of 4068 PoIs in Melbourne, after aggregatingall three aforementioned datasets. To crawl the opening hours forall those places, we use two Google Map APIs: Place Search andPlace Details. The former one provides a place_id for each placewhich is used for searching the details in the latter one. We crawlthe opening hours for POIs in both Melbourne and Rye, and weget a result involves totally 50 of POIs within eight different sub-categories.For a given parking lot at a specific date-time, we first calculatethe distance between all POIs and this parking lot, then we extractthe features based on the opening info retried in the former stage.We also record its minimum distance to an opening or any POI.After the extraction, we apply an ANOVA (Analysis of variance)test on both datasets. As shown in Table 3 and 4, the Pearson Corre-lation Coefficient of the same feature tends to have a contradicting oodstock ’18, June 03–05, 2018, Woodstock, NY anonymous author, et al. (a) Cluster on connection only. This is one small sample of park-ing slots in the Docklands area. Each rectangle represents aparking slot. Group 2 (red) and Group 7 (yellow) are close to eachother and have the same parking rule, while they are not in thesame group. (b) Cluster on distance and rules. If the distance of two park-ing slots within the given threshold and have the same park-ing rules, they are in the same group. The original Group 2 andGroup 7 are now in the same group (red), and group 11, 12 and13 are in the same group (purple).(c) The same area in Google Satellite Map
Figure 2: The Sample of Parking Slots Clusters in Melbourne CBD impact. Surprisingly, some contextual features has a opposite cor-relation with prediction occupancy in Melbourne CBD dataset andRye dataset. For example, humidity has a negative correlation withparking occupancy in Melbourne city but has a positive correlationin Rye area. This reflects the possible shift in the data distributionsof those two datasets, hence proving the need for introducing adomain transfer learning method in this situation.
The traditional method for transfer learning is fine-tuning, whichfirst loads a pre-trained parameter from other tasks and then re-train them on the new domain/task. However, one issue that needsto be faced in real-world usage is that most of the task only hasfew or no historical data at all. According to [4], Tzeng et al. [30]propose a general architecture for adversarial domain adaptionnamed Adversarial Discriminative Domain Adaption (ADDA). Thisnew framework used in ADDA combines a discriminative model,untie weight sharing and GAN loss together, which shows a promis-ing performance on unsupervised transfer learning. Compare toother domain adaptation methods, ADDA introduces the adversar-ial mechanism which trains an encoder to translate the featuresfrom the target domain to the latent space shared by both the source and target domain. Meanwhile, a discriminator is trainedsimultaneously to distinguish the origin of each latent code.In this paper, we adopt the original ADDA framework, which isinitially used for image classification task and modify it to makesit applicable for our time-series prediction problem. We use X s and X t to donate source and target domain features. Y s denotesoccupancy rate of parking lots from the source domain. M s ( X s ) donates source mapping/encoder and M t ( X t ) is about target map-ping. The regression model is represented as F while D stands forthe discriminator. The architecture we use in this paper is shownin Figure 3, and it comprises the three following stages.The first part is the pre-training step to learn a source encoder M s ( X s ) and a regression model based on the source domain data.Similar to an auto-encoder structure, the encoder here learns amapping the source domain to a latent space. On the other hand,the regressor learns to decode features from this latent space andmake a prediction on top of that. We use ConvLSTM (ConvolutionalLong-Short Term Memory) proposed in [33] as the encoder, whichshows a good performance on spatio-temporal data. Extending ona common LSTM unit, matrix multiplication is replaced by convo-lution operation at each gate in the LSTM cell. The key equationsof ConvLSTM are shown in equation 1 below, where ’ ∗ ’ denotesthe convolution operator and ’ ◦ ’ denotes the Hadamard product: ADACS: A Few-shot Adversarial Domain Adaptation Architecture for Context-Aware Parking Availability Sensing Woodstock ’18, June 03–05, 2018, Woodstock, NY
Source domain with historical data
Pre-training
Target domain without historical dataClassifer
Source convLSTM Source convLSTM D o m a i n D i sc r i m i na t o r Target convLSTMValue Domain label Target convLSTM
Classifer
Testing
Value
Adversarial Adaptation
Source domain with historical data Target domain without historical data
Figure 3: FADACS Domain Adaptation ArchitectureTable 2: The detailed description of the contextual featuresthat we used in this paper.
Feature Name Descriptionnum_of_open_poi 1.0 Number of Open POI within0.5 KM nearby.num_of_open_poi 0.5 Number of Open POI within1.0 KM nearby.Temp The temperature degree in Celsius.Hour The hour of the day.Wind The wind speed.num_of_poi 0.5 Number of POI within 0.5 KMnearby.Day Of Week The ordinal of the day in thewhole week.num_of_poi 1.0 Number of POI within 1.0 KMnearby.availability Whether this parking lot iscurrently available.Day Of Month The ordinal of the month in thewhole year.Barometer The barometer value.Extreme_weather An binary indicator for extremeweather.min_dis 1.0 The Shortest distance of POI nearbywithin 1.0 KMmin_dis 0.5 The Shortest distance of POI nearbywithin 0.5 KMHumidity The humidity value. i t = σ ( W xi ∗ X t + W hi ∗ H t − + W ci ◦ C t − + b i ) f t = σ ( W xf ∗ X t + W hf ∗ H t − + W cf ◦ C t − + b f ) C t = f t ◦ C t − + i t ◦ tanh ( W xc ∗ X t + W hc ∗ H t − + b c ) o t = σ ( W xo ∗ X t + W ho ∗ H t − + W co ◦ C t + b o ) H t = o t ◦ tanh ( C t ) (1) Next step is an adversarial adaptation, which is to learn a targetencoder M t ( X t ) so that the discriminator D cannot distinguish theorigin of that sample. By fixing source encoder parameter, the adver-sarial loss is used to minimise the distance of the mapping betweensource and target domain: M s ( X s ) and M t ( X t ) and maximise thediscriminator loss.min D L adv D ( X s , X t , M s , M t ) = − E x s ∼ X s [ log D ( M s ( X s ))]− E x t ∼ X t [ log ( − D ( M t ( X t )))] min M t L adv M ( X s , X t , D ) = − E x t ∼ X t [ log D ( M t ( X t ))] (2)In the final stage, we assemble the learned target encoder M t ( X t ) and regression model F together, and use data from the targetdomain to test its performance. The regressor should the ability togenerate quality prediction since the latent features from the targetdomain is overlapping with the ones from the source domain afterthe previous adaptation stage. We conduct all of our experiments on a Linux Server (CPU: IntelXeon Gold 6132 CPU @ 2.60GHz - 56 cores, GPU: NVIDIA QuadroGP100). In order to find the best parameters, we use a parallel gridsearch strategy that utilises all cores in this Linux cluster. As statedin the Data Pre-processing section, we use 5 minutes as the basicinterval between records. Besides, each sample contains featuresfrom the recent 30 minutes (i.e. 6 data points for each sample), andthe tasks are to predict the parking occupancy rate in the next 5,15 and 30 minutes (the next 1, 3, 6 timesteps). Two parking sensordatasets collected from Melbourne, Victoria and Rye, Victoria areused. The former one covers a whole year time period (2017), whilethe second one has a time range from 17th Nov 2019 to 20th Feb 2020.That reflects the big difference in both spatial and temporal domainwhich makes it difficult to apply the transfer learning method. oodstock ’18, June 03–05, 2018, Woodstock, NY anonymous author, et al.
Table 3: ANOVA test results of data from Melbourne City in Feb 2017
Features PearsonCorrelationCoefficient F Value p-Valuenum_of_open_poi 1.0 0.34 51742.63 0num_of_open_poi 0.5 0.33 48513.95 0Temp 0.23 21622.33 0Hour 0.16 9896.81 0Wind 0.13 6871.60 0num_of_poi 0.5 0.09 3356.62 0DayOfWeek 0.07 2063.57 0num_of_poi 1.0 0.05 1149.31 1.48e-251availability 0.02 108.92 1.70e-25DayOfMonth 0.01 78.61 7.60e-19Barometer -0.01 22.81 1.79e-06Extreme_weather -0.03 275.92 6.13e-62min_dis 1.0 -0.04 570.99 4.22e-126min_dis 0.5 -0.04 570.99 4.22e-126Humidity -0.25 26409.49 0
Table 4: ANOVA test results of data from Rye in Feb 2020
Features PearsonCorrelationCoefficient F Value p-ValueHumidity 0.19 9247.09 0DayOfMonth 0.11 3032.63 0Barometer 0.06 913.98 2.08e-200availability 0.02 118.98 1.07e-27num_of_poi 1.0 0.00 5.66 1.74e-02num_of_poi0.5 -0.03 245.34 2.86eDayOfWeek -0.06 771.68 1.41e-169Wind -0.07 1165.69 6.82e-255Hour -0.07 1139.34 3.41e-249min_dis 1.0 -0.12 3487.51 0min_dis 0.5 -0.12 3487.51 0Temp -0.24 14862.64 0num_of_open_poi 0.5 -0.30 24221.17 0num_of_open_poi 1.0 -0.32 27946.25 0Extreme_weather nan nan nan
In this paper, Mean Absolute Errors (MAE)and Root Mean Squared Errors (RMSE) are used to evaluate theeffectiveness of different models. Except for the adversarial adaptionstage, all models are trained using RMSE as its loss function.
For FADACS, we implement two variants: • ADDA (MLP): using MLP (Multi-Layer Perceptron) as theencoder to learn the mapping from the source/target domainto the latent space. • ADDA (ConvLSTM): using ConvLSTM as the encoder tolearn the mapping from the source/target domain to thelatent space. The intention here is to extract better latent features using ConvLSTM since the problem here is a spatial-temporal prediction problem.We compare FADACS with the following baselines: • HA (Historical Average): using the mean of historical dataas the prediction of the future data. • MLP: (Multilayer Perceptron): a feed-forward neural net-works which is widely used in function approximation andgeneral regression problems. It also relies on the feature ex-traction and is data hungry. It cannot distinguish temporalfeatures and spatial features. • LSTM (Long-Short Term Memory): a recurrent based methodthat is wide-used in many time-series prediction tasks [25].
ADACS: A Few-shot Adversarial Domain Adaptation Architecture for Context-Aware Parking Availability Sensing Woodstock ’18, June 03–05, 2018, Woodstock, NY
Table 5: Performance comparison with full parking data before domain adaptation
Model MAE (5/15/30 mins) RMSE (5/15/30 mins)HA 0.0600 0.1219MLP 0.0536 / 0.0895 / 0.1188 0.0988 / 0.1456 / 0.1771LSTM 0.0419 / 0.0767 / 0.1011 0.0942 / 0.1443 / 0.1765ConvLSTM
Model MAE (5/15/30 mins) RMSE (5/15/30 mins)ConvLSTM 0.0607 / 0.1091 / 0.1385 0.1222 / 0.1680 / 0.2003LSTM 0.0829 / 0.1035 / 0.1273 0.1261 / 0.1695 / 0.1998ADDA(MLP) 0.0845 / / 0.1774 0.1187 / / 0.2434FADACS / 0.1216 / 0.1694 / 0.1739 /
But it only focuses on the temporal domain. Therefore, if thespatial domain also plays an important role, its performancewill be limited. • ConvLSTM: a state-of-the-art methods used in transfer learn-ing area that can utilise features from both spatial and tem-poral domain [31].We conduct two sets of experiments based on the aforementionedbaselines. The first experiment is the basic parking occupancy pre-diction experiment. In the first experiment, all models are trainedand validated using data from the Rye dataset. This experimentmainly show the performance of existing method to parking predic-tion problem. For the transfer learning part, we apply our refinedADDA architecture on data from Melbourne and Rye to evaluate itsperformance. Namely, we choose the Melbourne data as the sourcedomain and the Rye data as the target domain since the Melbournedataset is much richer. Besides, we also train an LSTM model and aConvLSTM model on the source domain and test their performanceon the Rye data in this experiment.
In the first experiment, we compare a couple of existing approachesto predict the parking occupancy. We select four classic approachhere: HA, MLP, LSTM and ConvLSTM. HA is a basic statisticalmethod to estimate the parking occupancy based on the historicaldata by averaging them. The strength of this method is that HA cancatch the periodical pattern of parking occupancy. However, it doesnot consider spatial dependency, temporal dependency and hiddentrends in the data. Compared to HA, MLP can automatically explorethe trends of the parking occupancy even though it also does notconsider the spatio-temporal dependency. LSTM can predict theparking occupancy by leveraging the temporal dependency of thehistorical data which is the essential to time-series data prediction.However, as we mentioned in the introduction, the parking sens-ing not only relies on the temporal dependency but also relevantto the spatial dependency. ConvLSTM can integrate spatial andtemporal features into one simple end-to-end model and Table 5also validates our assumption. In Table 5, ConvLSTM outperforms other classic parking prediction approach for all prediction hori-zons. LSTM outperforms the second since it consider the temporaldependency but not spatial dependency. MLP performs better thanHA but lose the match to LSTM and ConvLSTM. This result sug-gests us that both spatial and temporal dependency play a role inthe parking occupancy prediction, and the temporal dependencyseems more important since the gap between the LSTM and MLPis much smaller than MLP and other approaches.The first experiment shows that ConvLSTM perform the best inparking sensing. Then, we conduct a few-shot transfer learning testto validate the effectiveness of our proposed transfer learning modelwith a few training samples from the target domain. most machinelearning techniques require thousands of examples to achieve goodperformance in parking prediction. The goal of few-shot learningis to achieve acceptable accuracy in parking sensing with a fewtraining examples in target domain. We compare our model to fourclassic approaches used in spatio-temporal transfer learning area:LSTM with parameter transfer, ConvLSTM with parameter transfer,ADDA with MLP and our propose architecture. The first and secondmodel are based on parameter transfer framework, which transferthe parameters trained in the source domain to the target domain.ADDA with MLP and our proposed architecture are GAN-basedtransfer learning framework. Table 6 shows that our approachperform the best. The ConvLSTM with parameter transfer performbetter than LSTM with parameter transfer, and the ADDA with MLPperform the worst. This result validates our claim that both spatialand temporal dependency are significantly important in parkingoccupancy prediction, and adversarial learning is a good at learningthe shared feature spaces. Additionally, it again validates that theimportance of each component should be temporal dependency,spatial dependency and domain adaption.In summary, we have conducted two experiments with Mel-bourne CBD parking data, Rye parking data and multiple contex-tual features. The experimental results show that our approachwhich integrates spatial information, temporal information anddomain adaption outperform other baselines. It also shows the im-portance of each component in predict parking occupancy in targetdomain by leveraging source domain historical data and contextualinformation. oodstock ’18, June 03–05, 2018, Woodstock, NY anonymous author, et al.
In this paper, we use both implicit sensing and explicit sensingapproaches to predict the parking occupancy in two different cities.We propose a GAN-based ConvLSTM transfer learning frameworkto sense the parking occupancy in a new area with few historicalparking data. We also qualitatively analyse the correlation betweenthe contextual information and parking occupancy with ten million-level real-world datasets. We compare our proposed model withthe state-of-the-art spatio-temporal transfer learning approach,and the experimental results show that our proposed model cansolve both significant challenges: spatial and temporal informationintegration and contextual information shared feature extraction.Our framework can be easily extend to other cities and other spatio-temporal sensors datasets as long as the data is graph-based andspatial correlated on which our model relies on. In the future, wewould like to investigate other data sources such as traffic andhuman mobile. We also will apply our proposed framework toother graph-based sensor data sensing problems.
REFERENCES [1] Andrew Arnold, Ramesh Nallapati, and William W Cohen. 2007. A comparativestudy of methods for transductive transfer learning. In
Seventh IEEE InternationalConference on Data Mining Workshops (ICDMW 2007) . IEEE, 77–82.[2] Rich Caruana. 1997. Multitask learning.
Machine learning
28, 1 (1997), 41–75.[3] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2018. Bert:Pre-training of deep bidirectional transformers for language understanding. arXivpreprint arXiv:1810.04805 (2018).[4] Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, HugoLarochelle, François Laviolette, Mario Marchand, and Victor Lempitsky. 2016.Domain-adversarial training of neural networks.
The Journal of Machine LearningResearch
17, 1 (2016), 2096–2030.[5] Taesik Gong, Yeonsu Kim, Jinwoo Shin, and Sung-Ju Lee. 2019. MetaSense: few-shot adaptation to untrained conditions in deep mobile sensing. In
Proceedings ofthe 17th Conference on Embedded Networked Sensor Systems . 110–123.[6] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley,Sherjil Ozair, Aaron Courville, and Yoshua Bengio. 2014. Generative adversarialnets. In
Advances in neural information processing systems . 2672–2680.[7] Bin Guo, Jing Li, Vincent W Zheng, Zhu Wang, and Zhiwen Yu. 2018. Citytransfer:Transferring inter-and intra-city knowledge for chain store site recommendationbased on multi-source urban data.
Proceedings of the ACM on Interactive, Mobile,Wearable and Ubiquitous Technologies
1, 4 (2018), 1–23.[8] Haviluddin Haviluddin and Rayner Alfred. 2014. Daily Network Traffic PredictionBased on Backpropagation Neural Network.
Australian Journal of Basic andApplied Sciences
International journal ofenvironmental research and public health
15, 4 (2018), 566.[10] Evangelia Kokolaki, Merkourios Karaliopoulos, and Ioannis Stavrakakis. 2013. Onthe efficiency of information-assisted search for parking space: A game-theoreticapproach. In
International Workshop on Self-Organizing Systems . Springer, 54–65.[11] Siyuan Lu, Zhihai Lu, and Yu-Dong Zhang. 2019. Pathological brain detectionbased on AlexNet and transfer learning.
Journal of computational science
IEEETransactions on knowledge and data engineering
22, 10 (2009), 1345–1359.[21] Chu Pengzi, Yang Jingshuai, Zhang Li, Gao Chong, and Sun Jian. 2017. ServiceData Analyze for the Available Parking Spaces in Different Car parks and TheirForecast Problem. In
Proceedings of the 2017 International Conference on Manage-ment Engineering, Software Engineering and Service Sciences - ICMSS ’17 . ACMPress, Wuhan, China, 85–89.[22] Christoph PflÃijgler, Thomas KÃűhn, Maximilian Schreieck, Manuel Wiesche,and Helmut Krcmar. 2016. Predicting the Availability of Parking Spaces withPublicly Available Data.[23] Zhou Qin, Zhihan Fang, Yunhuai Liu, Chang Tan, Wei Chang, and DeshengZhang. 2018. EXIMIUS: A measurement framework for explicit and impliciturban traffic sensing. In
Proceedings of the 16th ACM Conference on EmbeddedNetworked Sensor Systems . 1–14.[24] Alok Ranjan, Prasant Misra, Arunchandar Vasan, Sunil Krishnakumar, and AnandSivasubramaniam. 2019. City Scale Monitoring of On-Street Parking Violationswith StreetHAWK. In
Proceedings of the 6th ACM International Conference onSystems for Energy-Efficient Buildings, Cities, and Transportation . 31–40.[25] Wei Shao, Yu Zhang, Bin Guo, Kai Qin, Jeffrey Chan, and Flora D. Salim. 2019.Parking Availability Prediction with Long Short Term Memory Model. In
Green,Pervasive, and Cloud Computing
The high cost of free parking: Updated edition
The IEEE Conference on Computer Visionand Pattern Recognition (CVPR) .[31] Leye Wang, Xu Geng, Xiaojuan Ma, Feng Liu, and Qiang Yang. 2019. Cross-CityTransfer Learning for Deep Spatio-Temporal Prediction. In
Proceedings of theTwenty-Eighth International Joint Conference on Artificial Intelligence, IJCAI-19 .International Joint Conferences on Artificial Intelligence Organization, 1893–1899. https://doi.org/10.24963/ijcai.2019/262[32] Leye Wang, Bin Guo, and Qiang Yang. 2018. Smart city development with urbantransfer learning.
Computer
51, 12 (2018), 32–41.[33] Shi Xingjian, Zhourong Chen, Hao Wang, Dit-Yan Yeung, Wai-Kin Wong, andWang-chun Woo. 2015. Convolutional LSTM network: A machine learning ap-proach for precipitation nowcasting. In
Advances in neural information processingsystems . 802–810.[34] Shuguan Yang, Wei Ma, Xidong Pi, and Sean Qian. 2019. A deep learning ap-proach to real-time parking occupancy prediction in transportation networksincorporating multiple spatio-temporal data sources.
Transportation ResearchPart C: Emerging Technologies
107 (2019), 248–265.[35] Jason Yosinski, Jeff Clune, Yoshua Bengio, and Hod Lipson. 2014. How transfer-able are features in deep neural networks?. In
Advances in neural informationprocessing systems . 3320–3328.[36] Fengquan Yu, Jianhua Guo, Xiaobo Zhu, and Guogang Shi. 2015. Real time predic-tion of unoccupied parking space using time series model. In . IEEE, Wuhan, China,370–374. https://doi.org/10.1109/ICTIS.2015.7232145[37] Mingmin Zhao, Tao Ye, Ruipeng Gao, Fan Ye, Yizhou Wang, and Guojie Luo. 2015.Vetrack: Real time vehicle tracking in uninstrumented indoor environments. In