*3.5. Transferability*

Within the research context, it is common to employ test data to assess the models. Regardless if these data are obtained from real sources or synthetically generated, the resulting models have been built around them, and can be heavily linked to that experimentation context. Would these models work in other context or with other input data? Transferability could be defined as the quality of a data-driven model to be applied in other environment with other data, and it is directly linked with actionability: the application of a model should be generalizable to different datasets and transportation settings. This definition stems from the more general concept of *Transfer Learning* [156], which can entail that models trained in a certain domain are applied to other domains, so that the previous knowledge obtained from the first makes them perform better in the latter than models without it.

Depending on the subcategory of ITS, this requirement can be easily met or arduous to achieve, as some subcategories are more oriented towards the application and rely less on the environment than others; the key is defining what is *environment*. For example, a travel time forecasting model developed with data of a certain location could be transferable to another location without grea<sup>t</sup> complications, if it is built considering this feature [157]. In fact, many ITS models that are spatial-sensitive are developed using real data, but within the experimentation context, they are evaluated only in certain locations. Transferability for these scenarios would imply that the obtained results are reproducible (with certain degree of tolerance) in other locations.This could entail from plainly extrapolating the model to other locations [158], to implementing of techniques such as *soft-sensing*, aimed at modeling situations where no sensor is available [159], and the environment information is enough to obtain these models. A similar case in terms of spatial contexts, but with more parameter complexity, requires plenty of information about the environment. As an illustration, the case of crash risk estimation implies a higher calibration and adjustment needs due to the higher number of parameters that take part in this type of estimations. In these circumstances, works such as [160] or [161] work with posterior probability models and give more relevance to models that behave with a certain performance in many contexts than to models that perform better in a particular location. On the other extreme, for cases like autonomous driving, the change of environment is connatural to the domain (a moving vehicle constantly changing its location), and the parameters of these models are abundant and highly variable. Thus, these applications need transferable solutions, transferability that is specifically sought by researchers, for instance in LIDAR based localization [162] or pedestrian motion estimation [163]. In any case, and regardless the domain, ITS research is in an incipient stage (probably with the exception of autonomous driving) of developing transferable models, and evaluating this feature, and some machine learning paradigms can help improve this characteristic.

### **4. Emerging AI Areas towards Actionable ITS**

We have hitherto elaborated on the requisites that a model should meet towards leading to actionable data-based insights in ITS applications and processes. Some of these requirements can be fulfilled by properly designing the data-based workflow (e.g., interpretability can be straightforward for certain prediction models, whereas adaptability can be enforced by periodically scheduling the learning algorithm under use and feeding it with new data). However, several research areas have stemmed in the last years from the wide fields of Data Science and Artificial Intelligence that may serve to catalyze the compliance of data-based ITS workflows with the prescribed requisites, and thereby attain the sought actionability of their produced insights.

The main AI areas that have been identified as potentially appropriate for addressing the requirements can be summarized briefly as follows:

• Real-time data processing and online learning, which are not brand new research avenues in ITS, as we can find advanced developments in the literature. However, as we will later show, emerging fields with grea<sup>t</sup> potential such as dynamic data fusion and dynamic optimization can expedite and proliferate the adoption of incremental data-based models in more ITS-related applications.


We next discuss on the research opportunities spurred by the above research lines, their connections with the requirements presented in Section 3 (shown in Figure 3), as well as the challenges that stem from the consideration of these AI areas in the context of ITS.

**Figure 3.** Schematic diagram showing how avant-garde AI subareas can promote actionability in ITS data-based modeling workflows. Subareas contributing with particular emphasis to different functional requirements are connected together along the way from data to actions.

### *4.1. Online Learning and Dynamic Data Fusion/Optimization*

Previously sketched in Section 3.2, by online learning we refer to the capability of the learning model and in general, of the entire workflow, to learn from fastly arriving data possibly produced by non-stationary phenomena that enforces a need for adapting the knowledge captured by the model along time. Changes over data streams can make the

data pipeline obsolete, thus demanding active or passive techniques to update it with the characteristics of the stream [7,97].

Although activity around online learning has mostly revolved on certain clustering and classification paradigms (the latter giving rise to the so-called concept drift term to refer to pattern changes), it is important to note that adaptation can be also needed in other stages of the actionable data-based workflow, from data fusion to the prescription of actions. This being said, research areas such as dynamic optimization and dynamic multisensor data fusion should be also investigated deeply in future studies related to actionable data-based models, specially when the scenario under analysis can produce information with non-stationary statistical characteristics. When merging different data sources, fusion strategies at different levels can be designed and implemented, from traditional means (data-level fusion, knowledge-level fusion) to modern methods (corr. model-based fusion, federated learning or multiview learning) [164,165]. Fusion of correlated data sources can compensate for missing entries or noisy instances in static environments. However, when data evolve over time as a result of their non-stationarity, new challenges may arise in regards to the inconsistency among multiple information sources, including measurement discrepancy, inconsistent spatial and temporal resolutions, or the timeliness/obsolescence of the data flows to be merged, among other issues. For this reason, close attention should be paid to advances reported around adaptive fusion methods capable of detecting, counteracting and correcting misalignments between data flows that occur and evolve over time. This branch of dynamic data fusion schemes aims at combining together information flows produced by non-stationary sources, synthesizing a representation of the recent history of each of the flows to be merged into a set of more coherent, useful data inputs to the rest of the data-based pipeline [166,167]. On the other hand, dynamic optimization techniques can efficiently deliver optimized actionable policies when the objectives and/or constraints of the underlying optimization problem varies [168,169]. We energetically advocate for a widespread embrace of advances in these fields by the ITS community, emphasizing on those scenarios whose dynamic nature can make the obtained actionable insights eventually obsolete. This is the case, for instance, of traffic related modeling problems (e.g., traffic flow forecasting and optimal routing) or driver characterization for consumption minimization, among many others.

Other requirements for actionability can also benefit from the adoption of the above models in dynamic ITS contexts. For instance, cost efficiency in terms of energy consumption can largely harness the incrementality that often features an online learning model. The use of dynamic data fusion can also yield a drastically less usage of communication resources in wireless V2V links, such as those established in cooperative driving scenarios. All in all, the recent literature poses no question around the relevance of adaptation in databased modeling exercises noted in this work, with an increasing volume of contributions dealing with the extrapolation of adaptation mechanisms to ITS problems [170–172].

### *4.2. Transfer Learning and Domain Adaptation*

In close semantics to its related actionability requirement (*transferability*), transfer learning aims at deriving novel means to export the knowledge captured by a data-based model for a given task to another task with different inputs and/or outputs [156]. Depending on the amount of alikeness between the origin task and the destination task, we may be also referring to *domain adaptation*, by which we adapt the model built to perform a certain task to make it generalize better when processing new unseen inputs that do not follow the same distribution as their original counterparts (only the distribution changes [173]). Techniques such as subspace mapping, representation learning, of feature weighting arise as those methods most used to allow knowledge to be transferred between data-based models used for prediction.

In essence, transfer learning can provide higher prediction accuracy for models whose number of parameters to be learned (e.g., weights in a Neural Network) demands higher amounts of labeled data than those available in practice. However, data augmentation is not the only goal targeted by transfer learning. Domain adaptation may yield a better performance when used between ITS models that can become severely affected by a lack of calibration, different configurations or diverging specifications. An immediate example illustrating this hypothesis is the use of camera sensors for vehicular perception. Models trained to detect and identify objects in the surroundings of the vehicle can fail if the images provided as their inputs are produced by image sensors with new specs. The same holds for car engine prognosis: replaced components can make a data-based characterization of the normal operation of the engine be of no practical use unless a domain adaptation mechanism is applied. Personalization of ITS services can be another problem where domain adaptation can help refine a model trained with data from many sources: a clear example springs from naturalistic driving, where a behavioral characterization model built at first instance from driving data produced by many individuals (source domain) can be progressively specialized to the particular driver of the car where it is deployed [174–176].

In regards to actionability, several functional requisites can be approached by using elements from Transfer Learning over the data-based pipeline. To begin with, it should be clear that the transferability of learned models for their deployment in different locations and contexts could be vastly improved by Transfer Learning, as the purpose of this AI branch is indeed to meet this requirement in data-based learning models. In fact, this approach is currently under study and wide adoption within the ITS community working on vehicular perception: when the capability of the vehicle to sense and identify its surrounding hinges on learning models (e.g., Deep Learning for image segmentation with cameras), a plethora of contributions depart from pretrained models, which are later particularized for the problem/scenario at hand [177]. This exemplified use case supports our advocacy for further efforts to incorporate transfer learning methods in other ITS applications, specially those where data collection and supervision are not straightforward to achieve in practice. Another functional requirement where Transfer Learning can pose a difference in ITS developments to come is cost efficiency. The knowledge transferred between models learned from different contexts can improve their performance, thereby reducing the need for supervising data instances and ultimately, the time, costs and resources required to perform the data annotation.

Finally, the more recent paradigm coined as Federated Learning refers to the privacypreserving exchange of captured knowledge among models deployed in different contexts [178,179]. Although the main motivation for the initial inception of Federated Learning targeted the mobile sector, techniques supporting the federation of distributed databased models can be of utmost importance in the future of ITS, specially for V2V communications among autonomous vehicles and in-vehicle ATIS systems. Definitely the enrichment of models with global knowledge about the data-based task(s) at hand will pose a differential breakthrough in vehicular safety and driving experience. For instance, federated models can collectively identify, assess and countermeasure the risk of more complex vehicular scenarios than each of them in isolation [180]. Likewise, ATIS systems can learn from the preferences and habits of other users to better anticipate the preferences of the driver and act accordingly [181]. In a few words: an enhanced and more effective actionability of the data-based workflows built to undertake such tasks.

### *4.3. Gray-Box Modeling*

Gray-box modelling refers to the design of models that combine theoretical developments and structures related to the problem, with data that serve as a complement for such theories to make the overall model match better the scenario under analysis [182,183]. Gray-box models lie in between white-box models, for which the learned structure is deterministic and grounded in theoretical concepts; and black-box models, whose internal structure lacks physical significance and is learned from data. An example of white-box model in ITS systems is the use of computational fluid dynamics for macroscopic traffic flow modeling, whereas Deep Learning models for traffic forecasting can exemplify black-box modeling in this domain. Gray-box models have been lately embraced by the

ITS community in a number of modeling scenarios, such as those combining biological concepts and data-based models for driver characterization [184,185].

Gray-box modeling can contribute to the actionability of data-based workflows for ITS applications in two different albeit interconnected directions. To begin with, the incorporation of theoretical models to data-based pipelines can narrow the gap between engineers and practitioners more acquainted with traditional tools to analyze ITS systems and processes. Indeed, hybrid modeling can tie both worlds together not only without questioning the validity of prevalent theoretical developments, but also evincing the complementarity and synergy of both approaches. On the other hand, using validated theoretical models can help data-based modeling overcome difficult learning contexts such as class imbalance, outlier characterization or the partial interpretability of data clusters, among others.

### *4.4. Green Artificial Intelligence*

A profitable strand of literature has recently stressed on the energy efficiency of databased models, highlighting the need for redesigning their learning algorithms to minimize their energy consumption and thereby, make them implementable and usable in practice [186–188]. While this issue is particularly relevant for resource-constrained devices (e.g., mobile hand-helds), the concern with energy efficiency goes beyond usability towards environmental friendliness. For this reason many recent contributions are striving for computationally lightweight variants of machine learning models that sacrifice performance for a notable reduction of their energy demand. This is not only the case of predictive models capable of incrementally learning from data, but also of specific Deep Learning architectures tailored for their deployment on embedded devices [189].

Based on the above rationale, cost efficiency is arguably the most evident functional requirement around which energy-aware model designs can pose a breakthrough towards improving the actionability of the overall data-based workflow. In addition, other aspects can be made more actionable by using energy-aware model designs, such as usability [190]. Despite achieving unprecedented levels of predictive accuracy, a data-based workflow may become useless should it deplete the battery of the system on which it is deployed for operation. Therefore, energy efficiency should be under the target of future research efforts, specially when dealing with ITS applications running on battery-powered devices, inspecting interesting paths rooted thereon such as the trade-off between performance and energy consumption, or the adaptation of the model's operation regime depending on the remaining battery life, among others [191].

### *4.5. Fairness, Accountability, Transparency and Ethics*

To end with, the prescription of actions based on the insights provided by a data-based pipeline must be buttressed by a thorough understanding of the mechanisms behind its provided decisions [192]. Extended information about the model must be presented to the end user for several reasons:


• To supervise the ethics of data-based workflows, identifying potentially illegal uses of unlawful data given the prevailing legislation, guaranteeing the privacy and governance of personal data by third-party data-based ITS applications and processes, and certifying that the output of the model's output does not favor inequalities in terms of gender, religion, race or any other aspect alike.

The above requirements have been lately collectively compiled under the FATE (Fairness, Accountability, Transparency and Ethics) concept, which refers to the design of actionable data-based pipelines whose internal operations can be explained, accounted and critically examined in regards to the consequences of their eventual bias in privacy, fairness and ethical issues [193–195]. This recent concern with the operation of machine learning models spawns from the proliferation of real cases where practical model installments have unveiled deficiencies of different kind, from differential privacy breaks (data revealing the identity of the persons to whom they belong) to unnoticed output bias that caused racist discriminatory issues [196]. For instance, data-based models for vehicular perception, obstacle detection and avoidance must be also endowed with ethics and legal design factors to make the overall decision not just drifted by the data themselves. Another clear domain where FATE can be crucial is modeling with crowd-sourced Big Data, where aspects like privacy preservation [197] and bias avoidance [198] are arguably more critical [199,200]. The construction of the data-based modeling workflow must (i) ensure that protected features remain as such once the workflow has been built, without any chance for reverse engineering (via e.g., XAI techniques [90]) that could compromise the differential privacy of data; and (ii) that learning algorithms along the workflow counteract hidden bias in data that could eventually lead to discriminatory decisions (due to skewed samples, tainted annotation, limited data sizes or imbalanced data). From our perspective, these are among the most concerning challenges in the exploitation of Big Data in ITS, and the main source of motivation for a number of recent studies in areas related to data-driven transportation systems such as pedestrian detection [201], autonomous vehicles [202,203] or urban computing [204]. Bias-related issues can be identified by a proper analysis of the decisions made by the workflow, which in turn requires models to be accountable and transparent enough to thoroughly characterize their sensitivity to bias, and how inputs and outputs (decisions) correlate in regards to protected features. It is also remarkable to note that several proposals have been made to quantify fairness in machine learning pipelines, yielding useful metrics that account for the parity of models when processing groups of inputs [205,206]. Without these aspects being considered jointly with performance measures, data-based ITS developments in years to come are at the risk of being restricted to the academia playground [207].

### *4.6. Other AI Research Areas Connected to Actionability*

The above areas have been highlighted as the main propellers for model actionability in ITS systems. However, it is worthwhile to mention other research areas from the AI realm that can also help completing the chain from data to actions:


likeliness of the workflow to cause discriminatory issues as the ones exemplified above. The history of these AI areas in the ITS community has been going for years now [3]. However, we here emphasize the crucial role of these techniques beyond performance boosting: the techniques originally aimed to counteract the effects of class imbalance in the output of data-based models could be also leveraged to reflect legal impositions that not necessarily relate to the model's performance nor can they be inferred easily for the attributes within the data themselves. The lack of compliance of the model with fairness and ethics standards does not necessarily render a performance degradation observed at its output, nor can it be inferred easily from the available data.


•Finally, the ever-growing scales of ITS scenarios demand more research invested in scaling up learning algorithms in a computationally efficient manner [229]. Automated traffic, smart cities, mobility as a service constitute ITS scenarios where a plethora of information sources interact with each other. Definitely more efforts must be invested in aggregation strategies for data-based models learned from different interrelated data ecosystems, either in a distributed fashion (e.g., federated learning) or in a centralized system (correspondingly, Map-Reduce implementations of databased models, cloud-based architectures, etc). Computational aspects of large-scale implementations should be also under study due to their implications in terms of actionability, such as the latency of the system when prescribing decisions from data. This latter aspect can be a key for real-time ITS applications for which the gap from data to actions must be shortened to its minimum.

### **5. Concluding Notes and Outlook**

This work has built upon the overabundance of contributions within the ITS community dealing with performance-based comparisons among data-based models. Our claim is that, as in any other domain of application, data-based modeling should bridge the gap between data and actions, providing further value to the ITS application at hand than superior model performance statistics. It is our firm belief that the research community should embrace actionability as the primary design motto, with negligible performance improvements being left behind in favor of relevant aspects such as adaptability, usability, resiliency, scalability or efficiency.

To provide a solid rationale for our postulations, we have first presented a reference model for actionable data-based workflows, placing emphasis on the different phases that should be undertaken to translate data into actions of added value for the decision maker. Adaptation has been highlighted as a necessary albeit often neglected processing step in data-based modeling, which allows models to be effective when deployed on dynamic ITS environments with time-varying data sources. Next, our study has listed the main functional requirements that models along the reference model should meet to guarantee their actionability, followed by an overview of incipient research areas in Data Science and Artificial Intelligence that should progressively enter the ITS arena. Indeed, advances in XAI, Online Learning, Gray-box Modeling and Transfer Learning are currently investigated mostly from an application-agnostic perspective. Their undoubted connection to actionability makes them the core of a promising future for data-based modeling in ITS systems, processes and applications.

Other research areas related to Artificial Intelligence beyond those covered in our reflections will surely spawn further opportunities for actionability in ITS, provided that they fully embrace their ultimate goal: to effectively support decision making. Among them, the use of Automated Machine Learning (AutoML [230]) for tuning data-based models should not only optimize performance-based metrics (e.g., finding a model that attains maximum accuracy for image segmentation in vehicular perception cameras), but also comply with other objectives and constraints that closely link to actionability (e.g., robustness against adversarial attacks, or a lower epistemic uncertainty of the model induced in its output). Unless all such actionability constraints are regarded as design objectives and accounted for as such in the automated discovery of new data-based pipelines, any incursion of AutoML in ITS will be of no practical value. For this to occur, it is our belief that the confluence of multiple functional and non-functional requirements in this automated design process will pave the way towards the massive adoption of multi-objective optimization algorithms as a massive framework to infer and analyze all trade-offs existing among the design objectives.

Data-based modeling has brought a deep transformation to ITS. A vast amount of research works in the field are produced by data-based modeling specialists attracted by the profusion of available data, and with limited knowledge of transportation. Data-based models are getting progressively more complex, increasing the gap between research and

practice. This situation calls for a change of paradigm, to a one in which actionability requirements of models is desired by researchers, and practitioners are aware of the technologies available to provide it. Model actionability is a grea<sup>t</sup> whole that can act as an incentive to perform smaller steps towards its realization. It is probably unthinkable to develop, in a research environment, a data-based model that meets all proposed requirements. However, addressing some of the postulated requirements while developing a competing data-based ITS model will make it closer to actionability. There is, therefore, a long road to be travelled in ITS model actionability, with interesting avenues around the thorough understanding of models, and the adoption of emerging AI technologies to endow data-based workflows with the requirements needed to make them actionable in practice. As exposed in our study, there is a germinal interest in these research topics. Nevertheless, we foresee vast opportunities for future work when model actionability is set as a design priority.

On a closing note, we advocate for a new dawn of Data Science in the ITS domain, where advances in modeling performance concurrently emerge along with histories and reports about how such models have helped decision making in practical scenarios. Data mining has limited merit without actions prescribed from its outputs, always in compliance and close match with the specificities of its context.

**Author Contributions:** I.L.: Conceptualization, Methodology, Investigation, Writing—Original Draft, Writing—Review & Editing. J.J.S.-M.: Conceptualization, Methodology, Investigation, Validation, Writing—Original Draft, Funding acquisition. E.I.V.: Conceptualization, Methodology, Writing—Original Draft, Funding acquisition. J.D.S.: Conceptualization, Methodology, Investigation, Supervision, Writing— Review & Editing, Funding acquisition. All authors have read and agreed to the published version of the manuscript.

**Funding:** This work was supported in part by the Basque Government for its funding support through the EMAITEK program (3KIA, ref. KK-2020/00049). It has also received funding support from the Consolidated Research Group MATHMODE (IT1294-19) granted by the Department of Education of the Basque Government.

**Institutional Review Board Statement:** Not applicable.

**Informed Consent Statement:** Not applicable.

**Data Availability Statement:** Data sharing not applicable.

**Conflicts of Interest:** The authors declare no conflict of interest.
