Context-Aware Personal Route Recognition ˇ Oleksiy Mazhelis1,3 , Indr˙e Zliobait˙ e2,3 , and Mykola Pechenizkiy3 1

3

University of Jyv¨ askyl¨ a, Jyv¨ askyl¨ a, Finland [email protected] 2 Bournemouth University, Poole, UK [email protected] Eindhoven University of Technology, Eindhoven, the Netherlands [email protected]

Abstract. Personal route recognition is an important element of intelligent transportation systems. The results may be used for providing personal information about location-specific events, services, emergency or disaster situations, for location-specific advertising and more. Existing real-time route recognition systems often compare the current driving trajectory against the trajectories observed in past and select the most similar route as the most likely. The problem is that such systems are inaccurate in the beginning of a trip, as typically several different routes start at the same departure point (e.g. home). In such situations the beginnings of trajectories overlap and the trajectory alone is insufficient to recognize the route. This drawback limits the utilization of route prediction systems, since accurate predictions are needed as early as possible, not at the end of the trip. To solve this problem we incorporate external contextual information (e.g. time of the day) into route recognition from trajectory. We develop a technique to determine from the historical data how the probability of a route depends on contextual features and adjust (post-correct) the route recognition output accordingly. We evaluate the proposed context-aware route recognition approach using the data on driving behavior of twenty persons residing in Aalborg, Denmark, monitored over two months. The results confirm that utilizing contextual information in the proposed way improves the accuracy of route recognition, especially in cases when the historical routes highly overlap.

1

Introduction

Personal route recognition (prediction) is one of the essential elements of intelligent transportation systems (ITS) [17]. Accurate predictions enable providing personalized real-time information about location-specific events, implementing applications for warning about congested areas, emergency or disaster situations, or dispatching location-specific advertisements. In addition, accurate route predictions can help hybrid vehicles to optimize their charging phases leading to economy of fuel [6]. While the drivers can be asked to specify (e.g. type) their destinations or routes, this is likely to be perceived as invasive by most of the drivers, especially for regularly repeated routes [5]. Therefore, to make the ITS

2

O. Mazhelis et al.

unobtrusive, the routes should be predicted automatically, without the need for driver input. Route prediction has been a subject of a number of studies in different settings. Ziebart et al. [18] employed a Markov Decision Process for predicting the driver’s route given the destination. Liao et al. [8] utilized the Rao-Blackwellized particle filters to infer traveler’s position, mode of transportation, trip segment and goal from GPS traces. Recently, instance-based learning and string matching has been used for predicting the route of a driver [5,9,14]. In order to predict the route in real time, similarities between the current trajectory of the driver and the routes observed in past are calculated and the most similar route is output as the prediction. One of the major limitations of such instance based learning approaches is the latency of making the prediction, due to the need to accumulate some history of the trip before predictions become accurate. This happens as drivers typically have some frequent starting points (e.g. home) from which several routes begin. Evidently, whenever the beginnings of two different routes are identical, the similarity values will be high until the two routes diverge. As a result, some time needs to pass from the start of the trip before the route that is being traveled can be accurately predicted. This drawback limits the utilization of the route prediction systems, since accurate predictions are needed as early as possible, not at the end of the trip. This paper is aimed at improving the route prediction accuracy in the beginning of a trip, by utilizing contextual information. This contextual information captures external factors which may affect the choice of a route made by the driver. Earlier research has found that contextual factors, such as time-of-day or weather [18], influence the driver’s route preferences. In this paper, we introduce a technique for utilizing such contextual information in the process of predicting the driver’s route at the beginning of the trip. The main contribution of the paper is the technique describing how to incorporate this information into the route similarity assessment. We evaluate our approach on a publicly available dataset that represents a driving log of 20 drivers whose car movements were monitored over a period of several months. The results demonstrate that incorporating contextual information for adjusting the route similarity function leads to more accurate predictions at the beginning of a trip. The remainder of the paper is organized as follows. In Section 2 we present the problem formulation and the state-of-the-art nearest neighbor approach for the route recognition. In Section 3 we introduce our method for contextual adjustment. In Section 4 we report the results of the experimental evaluation. Related work is discussed in Section 5. Finally, Section 6 concludes the study.

2

Route Recognition using Instance-Based Learning

In this section, we define the settings and describe the state-of-the-art route recognition method that uses instance based learning.

Context-Aware Personal Route Recognition

2.1

3

The setting of the route recognition task

A trajectory of driver movements is modeled as a sequence of position fixes z1:n = (z1 , z2 , . . . , zn ). The sequence is ordered in time, where indices 1 . . . n are the time stamps. A route r is defined as a spatial projection of this trajectory onto the digital road network r = f (z1:n ). The route is represented as a sequence of turn-points r ≡ (e1 , . . . , em ), including the departure, the junctions or joints where the the driver turns from one road to another, and the destination. It is assumed that movements happen on-road only. A set of the routes observed for a driver in the past is denoted as DR . This set of routes can be updated with new routes as time passes. Two trajectories (and hence routes) are considered equal, if their departures and the destinations are the same, and all the turns along the trip are the same. The route recognition task is formulated as follows: given a set of past routes DR of a driver and a time ordered sequence of the position fixes z1:k for the current trip, identify the most likely route r ∈ DR . Obviously, the goal is to identify r correctly as early as possible, that is with k  n. This task assumes that the majority of routes are not unique, but reappear in the history of a driver. The assumption is in line with the findings of Froehlich and Krumm [5]. This setting does not restrict appearance of new routes in the reference route list. The representation of routes relies of turning points, which are not that common in rural areas; however, the selection of routes to be taken is also not that high in rural areas. Thus, the task of personal route prediction is considered to be more relevant to the urban environment. 2.2

Baseline route recognition approach

The current state-of-the-art instance based learning (IBL) approaches for route recognition [5, 9, 14] use instances from the reference set directly to classify the items in the test set. Whereas the conventional IBL [10] deals with static instances, the route recognition task poses specific challenges: the length of the current instances is increasing when the trip progresses; furthermore, the turnpoints comprising the routes are ordered in time. The longest common subsequence (LCSS) metric [3, 16] does not require the instances to be of the same length, thus it is used for calculating the distances, similarly to [9]: LCSS(ri , r∗ ) = |(c1 , . . . , cm )|,

(1)

where ri = (e1 , . . . , en ) is the i-th reference route, r∗ = (e∗1 , . . . , e∗k ) is the current route, || denotes the length of a route, and (c1 , . . . , cm ) is the longest sequence of common turn-points such that ∀ch , h ∈ {1, . . . , m} ∃eih , e∗jh : ch ≡ eik , ch ∼ e∗jk , ∀h > 1 : ih−1 < ih ≤ ih−1 + δ, jh−1 < jh ≤ jh−1 + δ, where δ is a maximum mismatch threshold, and the turn-points ch and e∗jk are treated as equal (ch ∼ e∗jk ), if they have the same street code, and if they are physically situated within a predefined distance.

4

O. Mazhelis et al.

The LCSS allows parts of the two sequences to be different, which is a desired property for comparing trajectories. As a result, outliers in the position fixes and map-matching errors, which are likely due to noisy nature of the positioning data, can be tolerated. On the other hand, the LCSS is robust in discriminating between close, but diverging routes. The measure is based on the longest segment that is present in both compared trajectories, thus, even a minor divergence of the routes going beyond the predefined outlier threshold (e.g. driving a parallel street to the reference root) greatly reduces the value of LCSS and hence makes the routes distinguishable. The online route recognition is performed as follows. As the vehicle is moving, its position is periodically measured. The incoming position fixes are used to determine the location on the road, as well as identify turn-points along the route. The current instance r∗ for prediction is formed from a list of the preceding turn-points up to and including the latest turn-point e1:k . Every time a new turnpoint incomes and the current route is updated, the current instance r∗ changes and the prediction is updated. For every reference route in the database ri ∈ RD , its similarity with the current instance r∗ is calculated as ( LCSS(ri , r∗ )/R, if |r∗ | ≤ |ri |, Sim(ri , r∗ ) = 0 otherwise, where the module denotes the length of the route. In [9], LCSS has been normalized by the length of the traveled route, i.e. R = |r∗ |. In this study we add an additional component k1 |ri− | to the correction factor in order to put more emphasis to the similarity of the routes which have different starts but converge later, thus helping to predict the next points in the converged route similarly to the case when the routes start at a common point. Here ri− is the part of the reference route ri that lies prior to the common subsequence and k1 > 0 is a weight parameter. Following the same reasoning we amplify the common subsequences that are situated closer to the end of the current route. For that purpose we introduce a correction k2 |r∗+ |, where r∗+ is the part of the current route r∗ that follows the common subsequence and k2 > 0 is a weight parameter. Thus, in this study we use the following normalization for calculating the similarity between the current and a reference route: R = |r∗ | + k1 |ri− | + k2 |r∗+ |. Sim(ri , r∗ ) produces values between 0 and 1. Using the produced similarity values, the most similar route is nominated as ri : i = max Sim(ri , r∗ ). i

To avoid mistaking short stops for the end of the route, a route is assumed to be completed whenever the time spent in the same place exceeds tb . After completing each route, the system stores this route as an instance in the reference database DR . Only the instances with maximum similarity below smax are added to the reference set to avoid overloading. The reference set can be further truncated based on the time elapsed since the last recognition made.

Context-Aware Personal Route Recognition

3

5

Using Context Information for Route Recognition

Having discussed in the previous section the baseline IBL approach for route prediction, in this section we present our approach for incorporating contextual information into predictions. Route recognition is performed in real time, as the vehicle moves. If multiple routes have equal beginning, their similarity is likely to remain high until the routes diverge. In such cases, similarity alone is insufficient to decide which of the routes will be taken, and a number of routes that have the maximum similarity to the current trip may be nominated, as they are equally likely. We propose to alleviate this problem by incorporating into predictions contextual information, such as the time of the day or weather. We formulate two approaches: contextual selection and contextual correction. Contextual selection filters the reference routes leaving only those which have the same context as the currently observed instance. For example, if currently it is raining, only the routes that were completed when it was raining, will be left in the reference set. Contextual correction uses all the routes in the reference set; it adjusts the similarity scores based on the context. The similarity is increased to those routes that are in the same context as the current instance, and similarity is decreased to those routes that are in a different context. For example, if currently it is raining, all the routes completed when it was raining will get ‘bonuses’ towards their similarities with the current trip, but all the reference routes would remain in the reference set. This way, if there is an exact or very close match among the routes that were completed when it was not raining, this route may be picked up. Contextual selection is a simple version of taking context into account; it is expected to work well if the routes in different contexts are highly different. Contextual correction is expected to work if the routes are very similar, but the prior probabilities of taking a route in different contexts are different. The rest of the section describes how to learn the contextual correction. We propose to incorporate context into calculating the similarity Sim(ri , r∗ ) between the current route and each of the reference routes. We model the choice of routes taking into account the impact of context to the driver’s choice. The parameters of this model need to be learned on the reference dataset RD of the routes observed in the past for the same driver. After that, the learned model can be applied for prediction in real time. Given a training set consisting of n trips of different length we aim at learning how to adjust the similarity depending on the observed context. Let s∗ij = Sim(ri , rj ) denote the similarity between completed routes ri and rj . Let sijt = Sim(rit , rjt ) denote the similarity between parts of these routes from the start to time t. t can be measured in seconds, in trip distance, in turnings or in percentage thereof. We propose an additive correction for sijt at time t: sˆijt = sijt + ρij αt ,

(2)

where ρij = 1 if the context of the routes ri and rj is the same, otherwise ρij = −1; and αt is the dynamic correction coefficient, computed for each t. Note that we consider context as a property of a route, not a property of a time point. Thus a given route will have the same context all the way.

6

O. Mazhelis et al.

We form the following square loss function at time t: Lt =

n n X X

(sijt + ρij αt − s∗ij )2 .

(3)

j=1 i=1,i6=j

Let us find the value of αt minimizing Lt : n n X X ∂Lt = ∂αt j=1


2ρ2ij αt + 2ρij (sijt − s∗ij ) = 0.

(4)

i=1,i6=j

From here, the value of αt follows Pn Pn αt = −

j=1

i=1,i6=j

ρij (sijt − s∗ij )

n(n − 1)

.

(5)

We derived an expression for αt that minimizes the loss function at a particular time t. However, re-estimating αt at every time step t may be computationally inefficient and infeasible given limited historical datasets. Thus, following our result for time t, we construct an approximation for αt as a function of time. We model αt as a decreasing geometric progression, since we observe that the context information is the most important at the beginning of the trip, when little information about the trajectory is available. As the route is traveled further, the impact of context information diminishes, since the routes are likely to diverge or merge, and the trajectory similarity may be sufficient to correctly recognize the route. As argued, we model αt as αt = aq −t ,

(6)

where q > 1 is a fixed constant, and a is the parameter to be found. Our square loss function now is a sum of differences between the exact correction factor αt that comes from the data and the approximations of that correction over time 2 T  X a − αt . (7) L= qt t=1 This loss function is minimized when T

∂L X 2 = ∂a qt t=1



a − αt qt

 = 0.

(8)

We find the value of a by solving Eq. (8) PT  T T T  −t X X X 1 a αt αt t=1 αt q = 0; a − = ; a = . P T −2t q 2t qt q 2t qt t=1 t=1 t=1 t=1 q Here T limits the time-span, within which the context correction is applied. It is assumed that the correction is important mostly in the beginning of the trip, until sufficient information about the trajectory is available, and therefore a small value (e.g. T = 10 turn-points) is to be used.

Context-Aware Personal Route Recognition

BL: baseline

Trajectory information

Baseline route recognition

CC: context correction

Trajectory information

Baseline route recognition

7

Prediction

Context correction

Prediction

Context information CS: context selection

Trajectory information

Context selection

Baseline route recognition

Prediction

Context information

Fig. 1. Alternative route prediction approaches being compared

4

Performance Evaluation

The effect of the context-based correction on route prediction has been empirically evaluated using a real world database containing 20 datasets from different drivers recording in total 2346 trips. In this section, the experimental settings are described, and the results of the experiments are reported and discussed.

4.1

Experimental goals

The experiments are aimed at exploring the role of context in the driving behavior data and evaluating the effect of the proposed contextual correction on the final prediction accuracy. In the experiments we compare the accuracy of route prediction of the following three approaches (see Fig. 1). 1. The baseline (BL) is the state-of-the-art IBL approach described in Section 2, which does not take context information into account. 2. The context correction (CC) approach, where the similarity values are corrected as described in Section 3. 3. The context selection (CS) approach, where the candidate routes are prefiltered so that only the routes that have the same context as the current trip are left in the reference set. After that, the baseline approach is applied to the set of pre-selected routes. We expect context selection to work in cases when the routes in different contexts are different. We expect our contextual correction to work well when the routes in different contexts are generally the same, but the prior probabilities of choosing these routes are different in different contexts.

8

4.2

O. Mazhelis et al.

Data

The data used in the experiments have been gathered and made available by the INFATI project [7]. 20 cars and families participated in the data-gathering, and their daily driving was monitored over the period of two months in 2000-2001. The cars were driven in the municipality of Aalborg, Denmark, which includes the city of Aalborg, its suburbs, and some neighboring towns. Whenever a car was moving, its GPS position was sampled with 1Hz frequency, and stored in the UTM 32 format. These GPS fixes were also mapped onto the roads. In order to protect the privacy of the drivers, the log entries with GPS position in the vicinity of the residence of the driver were removed. The exact procedure for entry removal is described in [7]. For the purposes of the experiments, the INFATI data have been converted from the UTM 32 into the WGS84 reference coordinate system. Only the entries that were successfully mapped onto streets were used (the others were assumed to be outliers). Furthermore, for the sake of simplicity, whenever multiple drivers were driving a car, only the entries for one of the drivers (the one with the largest number of entries) were used. 4.3

Data preprocessing and route labeling

The original data does not contain information about the start and the end of a route and the routes are not labeled. For evaluation purposes we converted each dataset into a set of routes by splitting the sequences whenever the gap between two consecutive entries exceeded the threshold tb of five minutes. The first entry marked the end of one route and the second entry marked the start of the following route. Our threshold value tb is similar to the gap used in [5], and its appropriateness was verified empirically. For evaluation purposes, we obtained the ground true labels of the routes in the following way. At each time point t every pair of routes in a dataset was labeled either as ‘equal’ or ‘not equal’. Two routes were considered equal if their remaining parts from time t to the end were equal. This way two merging routes are treated as distinct (with different labels) until the merger point, and are treated as equal afterwards. We used the time of the day as a contextual feature. We split the day into five slots representing five contexts: 0-5 a.m., 5-10 a.m., 10 a.m.-3 p.m., 3 p.m.-8 p.m. and 8 p.m.-0 a.m. Each route was associated with one of these five slots based on when the trip started. 4.4

Evaluation criteria

We use a simple accuracy [12] as the primary measure of the performance. Note that because of using the dynamic labeling of routes in DR in the evaluation of route prediction at each time t we can have multiple routes in DR (excluding r∗ ) that have the same label as r∗ . We deal with this in the following way. For one driver the accuracy of predicting an individual route at time t is calculated as

Context-Aware Personal Route Recognition

9

At = (nTPt + nTNt )/|DR |, where TP and TN designate true positives and true negatives respectively. A true positive is registered if the solely predicted route is correct (i.e. has the same label as the current one); otherwise, a true negative is registered. However, often several routes have the maximum similarity, with n+ of them being correct + (denote the set thereof as DR ), while the other n− being incorrect (denote the − set thereof as DR ). Since a single route needs to be predicted, one of the routes having maximum similarities can be selected randomly, with probability of being correct p = n+ /(n+ + n− ). This probabilistic choice is taken into account when counting nTPt and nTNt , by assigning weight p to each of the routes having maximum similarity. Let mt be the max similarity at time t: mt = max Sim(rt∗ , rt ) rt ∈DR

When multiple routes have max domly. Then,  p,    p, At (r∗ , ri ) =  1,    0,

similarities, one of them can be selected ran+ Sim(rt∗ , rti ) = mt , rti ∈ DR t − Sim(rt∗ , rti ) = mt , rti ∈ DR t − Sim(rt∗ , rti ) < mt , rti ∈ DR t otherwise.

The first row in the definition above represents true positives; the next two represent true negatives. Accuracy at time t is estimated by averaging over all candidate routes: 1 1 X − − At (r∗ , ri ) = (pn+ At (r∗ ) = t + pnt + ut ), |DR | |DR | DR

− ∗ where u− t = |{rtj }| such that Sim(rt , rtj ) < mt , rtj ∈ DRt . Then, accuracy can be expressed as:   1 n+ 1 t − + − − At (r∗ ) = (n + n ) + u = (n+ t + ut ). t t t − |DR | n+ |D | + n R t t

4.5

Experimental protocol

We used the data from each driver as an independent dataset. We obtained the testing accuracy on each dataset using the leave-one-out cross-validation. Namely, every route of a driver played the role of the route to be predicted, while the remaining ones comprised the reference dataset DR . The learning and the testing phases are implemented as follows. In the learning phase, the values of αt are estimated based on Eq. 5. After that, the value of a is evaluated according to Eq. 8. In the testing phase, the value of αt is approximated based on Eq. 6. After that, the similarity values are corrected based on Eq. 2.

10

O. Mazhelis et al.

We use the contextual correction to adjust the similarity scores only in the beginning of the trip, where the trajectory alone is insufficient to distinguish among routes. The value of T defines the period of time, during which the adjustment is applied. The larger the T , the longer time the adjustment is performed. On the other hand, T also defines the minimum length of the routes applicable for learning αt , and the larger the T, the smaller number of routes could be used for estimating αt , thus potentially impeding the accuracy of the estimate. In order to explore the effect of T on the adjusted similarity function, the accuracy was estimated for three values of T : T = 7, T = 10, and T = 15.

4.6

Setting the parameters

Our approach for context correction is not parameter free, i.e. it requires setting the values of several parameters a priori. The details of a priori parameter setting are as follows. In Eq. (2) the coefficients k1 and k2 are aimed at decreasing the similarity measure if the LCSS is distant from the beginning of the reference route (k1 ) or from the end of the current route (k2 ). We assume that equal endings of the routes should contribute more to the similarity than equal beginnings (i.e. k1 < k2 ). Furthermore, we assume that the length of LCSS should have greater effect on the similarity measure than the (in)equality of the beginnings and endings of the routes, i.e. k1 < k2 < 1. Following this reasoning we fix the k1 and k2 values as k1 = 0.2 and k2 = 0.5. The value of q > 1 determines how quickly the estimate of a decreases with time (cf. Eq. 4). We fixed q = 1.2 based on visual inspection of the geometric progression plots with synthetic data; a small value was selected to achieve a flat decrease in the impact of context. To avoid any possible problem of overfitting, we fixed those parameters in the preliminary experiments with one driver (driver 14) and based on our reasoning.

4.7

Context in the dataset

Before discussing the results let us first consider how the routes depend on the context in our dataset. Figure 2 plots the routes of one driver in three different contexts. For illustration purposes these plots only connect the start (circle) and the end (triangle) of the routes with a straight line on the real coordinate axes. The figure shows that there are many overlapping routes, but the travel patterns are different in different contexts. For instance, in the early morning (5-10 a.m.) only specific routes are likely, probably to work. If all the traveling patterns look like this then we expect contextual selection to work. However, if we look at the midday (10 a.m.-3 p.m.) and the evening (3 p.m.-8 p.m.) the routes highly overlap. The arrow points a specific route which happens in both contexts, but is more likely in the evening. In this case we expect contextual correction to work.

Context-Aware Personal Route Recognition

y coordinate

10 a.m. − 3 p.m.

y coordinate

3 p.m. − 8 p.m.

y coordinate

5 − 10 a.m.

11

x coordinate

x coordinate

x coordinate

Fig. 2. Routes of one driver. 0.98

0.98

0.96

0.96

0.94

0.94

0.92

0.92

0.96 0.94 0.92

0.9

0.9

0.88

0.88

Acc

Acc

Acc

0.9 0.88 0.86 0.84 0.82 Baseline Context selection With correction

0.86 0.84

1

2

3

4

t

5

6

Baseline Context selection With correction

0.86

7

0.84 0

2

4

6

t

8

Baseline Context selection With correction

0.8 10

0.78 0

5

10

15

t

Fig. 3. Prediction accuracy for the ninth driver (T = 7, T = 10, and T = 15)

4.8

Effects of contextual correction to the route recognition accuracy

We ran our experiments on 20 datasets from 20 drivers using the time of the day as context feature. In this section we report route prediction performance for selected individual drivers; the next section reports the aggregated results. The approach with contextual correction outperformed the other two approaches for the majority of drivers. Figure 3 plots the route prediction accuracy obtained for one of the drivers, which we find to be one of the most illustrative. As could be seen, for this driver, the context correction gives the most accurate predictions, while the worst accuracy is provided by the baseline approach. However, there were individual cases where the use of context-based correction brought little or no improvement in accuracy. As an example, see Figure 4, which illustrates the performance on the data for one such driver given T = 15. The context selection approach outperforms the context correction. In the latter case the inferior accuracy of the contextual correction can be partly attributed to the way, in which merging routes are labeled. Two routes that merge at some point get different labels prior to the merger point, and equal labels afterwards. Thus, the similarity of these two routes is expected to be zero prior to the merger, and to start increasing afterwards. Recall that our contextual correction is trained to adjust the similarity towards the final value. Hence, the similarities of the turn-points preceding the merger are increasing

12

O. Mazhelis et al. 1

1

1

0.95

0.95

0.9

0.9

0.9

0.8

Acc

Acc

Acc

0.85 0.85

0.8

0.8

0.7 0.6

0.75 0.75

0.65

0.5

0.7 Baseline Context selection With correction

0.7

1

2

3

4

5

6

Baseline Context selection With correction

0.65 7

0

2

4

t

6

8

Baseline Context selection With correction

0.4

0

10

5

10

15

t

t

Fig. 4. Prediction accuracy for the first driver (T = 7, T = 10, and T = 15 respectively) Routes with 7 turn-points

Routes with 10 turn-points 1

1.5

Baseline Context selection With correction

Rank

Rank

1.5

2

2.5

3

Routes with 15 turn-points

1

1.5

Baseline Context selection With correction

Rank

1

2

2.5

1

2

3

4

t

5

6

7

3

Baseline Context selection With correction

2

2.5

2

4

t

6

8

10

3

2

4

6

8

t

10

12 14

Fig. 5. Averaged ranks of three different approaches

while labels stay different. As a result, the adjusted (increased) similarity may reach its maximum before the merge point, while the labels will still be different. If that happens, we observe a decrease in accuracy. 4.9

Aggregated results

To evaluate the overall performance of our proposed contextual correction we ranked the accuracies achieved with three alternative approaches at each time t and average the resulting ranks. Rank 1 indicates the top-most accuracy, rank 3 indicates the worst accuracy. Figure 5 plots the averaged ranks over 20 drivers. From the figure we observe that the accuracy of the context correction approach is systematically ranked as top-most, independently of time t and the limiting factor T . These results indicate that the suggested context-correction approach provides superior route prediction accuracy, as compared both with the context-free baseline approach and with the context-selection approach.

5

Related Work

Context-awareness is widely used in ubiquitous and pervasive computing to characterize the environmental variables [13]. In machine learning the term usually

Context-Aware Personal Route Recognition

13

characterizes the features that do not determine or influence the class of an object directly [4,15]. The strategies are not limited to building a separate predictor or a combination for each context. Contextual information can be also used to adjust the input data, model parameters or model outputs. Context handling strategies in recommender systems are grouped [1] into pre-filtering, post-filtering and contextual modeling. In pre-filtering, context management actions are taken before building actual classifiers. In post-filtering, classifier outputs are adjusted based on contexts. In contextual modeling, the context information is an explicit part of classifier training. Our contextual correction falls into a category of post-filtering approaches [1]. A model is learned first, and then the model outputs are corrected using context information. A few illustrative examples on how to train post-filtering component (in the domain of recommender systems) can be found in [2, 11]. The contextual selection can be seen as a pre-filtering. Contextual modeling is hardly suitable in route recognition tasks as our instances are of different length and they cannot be expanded with contextual information in a straightforward and uniform way. Contextual information has been used in other problems and applications, while to the best of our knowledge there is no directly related work for route recognition problem. The problem is specific, since the instances (trajectories) contain a time dimension and change over time. Similarly to our work, Ziebart et al. [18] utilize context information for route prediction; however, they formulate the task of route prediction differently. Namely, the authors assume that the destination is known, and attempt to predict the route to the known destination. Due to this formulation, the problem of distinguishing among routes having common start but leading to distinct destinations is avoided. Thus, Ziebart’s et al. work is aimed at assisting the drivers in selecting the route to a known destination, whereas our approach assumes that the route is already chosen by the driver, and strives to automatically recognize that route in order to supply additional information to the driver. We are not aware of any other works that would incorporate contextual information into real-time route recognition.

6

Conclusions and Future Work

The information about personal routes of drivers is critical for many ITS applications, varying from location-dependent advertisements to emergency warning applications. To increase the utility of such applications, these routes should be recognized in advance, preferably in the very beginning of the trip. If machine learning approaches rely only on GPS signals (even aligned to a known road map), timely recognition of a personal route appears to be a difficult task simply because often different routes start from the same departure point. In this paper, a novel approach to route recognition has been introduced. Our approach utilizes the information about the context of the trip aiming to improve the recognition accuracy at the beginning of the trips. More specifically, the context information is employed when assessing the route similarities,

14

O. Mazhelis et al.

by adjusting positively (negatively) the similarities of the routes with matching (mismatching) context. The degree of adjustment decays as the trip progresses, since the trajectory information being accumulated becomes sufficient for accurate route recognition. Both the degree of adjustment and the decay factor are learned on a training set, so as to minimize the route recognition error. The proposed approach has been empirically tested using the data obtained from 20 drivers whose driving behavior was monitored over the period of several months. In the experiments, the proposed approach has been compared with a simple approach where the context is used to pre-select the relevant routes, and against the baseline approach that does not take the context into account. The results of the experiments indicate that the proposed approach outperforms the baselines in terms of accuracy of predictions, especially at the beginning of a trip. There are several directions to extend the proposed approach for integrating contextual information into prediction of a personal route. First, the contextbased similarity correction assumes that the relationships between context and the routes remains unchanged. If this assumption does not hold, i.e. if the preferences of the driver in different context change, the context-dependency should be re-learned for the proposed approach to be effective. Detection of such changes and updating the context-based correction accordingly has been left for further study. Second, in the experiments, only the time of the day was explored as the context of driving behavior. Other contextual characteristics, such as traffic congestions, weekly and daily variations, weather conditions, road quality, etc. may affect the choice of the route. Whether the proposed approach is able to capture the regularities potentially present in these characteristics remains to be studied as a part of future work. The further work may also focus on the selection and optimization of correction parameters for an individual driver or for a category of similar drivers, as well as on similarity-based clustering of the routes for finding the most representative ones.

7

Acknowledgements

This research is partly supported by the SDFA project (a part of the Cooperative Traffic research program funded by the National Technology Agency of Finland) and LOIS visitor grant. The authors are thankful to the INFATI project for making the datasets used in this study available for the academic research.

References 1. G. Adomavicius and A. Tuzhilin. Context-aware recommender systems. In Proc. of the ACM Conf. on Recommender systems (RecSys’08), pages 335–336, 2008. 2. G. Adomavicius and A. Tuzhilin. Recommender Systems Handbook: A Complete Guide for Research Scientists and Practitioners, chapter Context-Aware Recommender Systems, page to appear. Springer, 2010.

Context-Aware Personal Route Recognition

15

3. L. Bergroth, H. Hakonen, and T. Raita. A survey of longest common subsequence algorithms. In Proceedings of the Seventh International Symposium on String Processing Information Retrieval (SPIRE 2000), pages 39–48. IEEE Computer Society, 2000. 4. P. Br´ezillon. Context in problem solving: a survey. Knowledge Engineering Review, 14(1):47–80, 1999. 5. J. Froehlich and J. Krumm. Route prediction from trip observations. In Society of Automotive Engineers (SAE) 2008 World Congress, 2008. 6. J. D. Gonder. Route-Based Control of Hybrid Electric Vehicles. In SAE 2008 World Congress. SAE International, March 2008. 7. C. S. Jensen, H. Lahrmann, S. Pakalnis, and J. Rung. The INFATI data. Timecenter technical report tr-79, July 2004. 8. L. Liao, D. J. Patterson, D. Fox, and H. Kautz. Learning and inferring transportation routines. Artif. Intell., 171(5-6):311–331, 2007. 9. O. Mazhelis. Real-time recognition of personal routes using instance-based learning. In 2011 IEEE Intelligent Vehicles Symposium, page to appear, 2011. 10. T. M. Mitchell. Machine Learning. McGraw-Hill Science/Engineering/Math, 1 edition, Mar. 1997. 11. U. Panniello, A. Tuzhilin, M. Gorgoglione, C. Palmisano, and A. Pedone. Experimental comparison of pre- vs. post-filtering approaches in context-aware recommender systems. In RecSys ’09: Proceedings of the third ACM conference on Recommender systems, pages 265–268. ACM, 2009. 12. M. Sokolova, N. Japkowicz, and S. Szpakowicz. Beyond accuracy, F-score and ROC: A family of discriminant measures for performance evaluation. In AI 2006: Advances in Artificial Intelligence, volume 4304/2006 of Lecture Notes in Computer Science, pages 1015–1021, 2006. 13. T. Strang and C. Linnhoff-Popien. A context modeling survey. In Workshop on Advanced Context Modelling, Reasoning and Management at the 6th Int. Conf. on Ubiquitous Computing (UbiComp 2004), 2004. 14. K. Torkkola, K. Zhang, H. Li, H. Zhang, C. Schreiner, and M. Gardner. Traffic advisories based on route prediction. In Proceedings of Workshop on Mobile Interaction with the Real World, pages 33–36, 2007. 15. P. Turney. The identification of context-sensitive features: A formal definition of context for concept learning. In Proc. of the ICML’96 Workshop on Learning in Context-Sensitive Domains, pages 53–59, 1996. 16. M. Vlachos, G. Kollios, and D. Gunopulos. Discovering similar multidimensional trajectories. In Proceedings of the 18th Int. Conference on Data Engineering, pages 673–684, 2002. 17. T. H. N. Vu, K. H. Ryu, and N. Park. A method for predicting future location of mobile user for location-based services system. Comput. Ind. Eng., 57(1):91–105, 2009. 18. B. D. Ziebart, A. L. Maas, A. K. Dey, and J. A. Bagnell. Navigate like a cabbie: probabilistic reasoning from observed context-aware behavior. In UbiComp ’08: Proceedings of the 10th international conference on Ubiquitous computing, pages 322–331, 2008.