Analyzing the tactical behavior in team sports is of paramount importance in sports performance analysis. The individual actions performed are of interest when analyzing the team's tactics. For quite some time, action frequencies of teams, and players have been the only way to gain insight into this performance aspect. However, this is not enough to get a complete picture of the performance, and especially the tactical behavior. Therefore, action/event sequences have been suggested for deeper insight into the game. One reason mentioned by Carling et al. (2008) is that “on-the-ball” activity, physical contact, and the sequence in which these actions occur contribute to physiological energy expenditure. This means that in addition to tactics, this kind of action analysis could give insight into player fatigue. Action sequences are chains of sequential single actions during a game (Schrapf et al., 2017). As the OPTA data are based entirely on event or action data, with timestamps and positional coordinates available, it is especially suitable for this type of analysis.

The paper presents a novel technique for sequence analysis of football event data and discusses its advantages and disadvantages for decision making in football. To the best of our knowledge, process mining has not been used for sports performance analysis until now.

Process mining aims at discovering, monitoring, and improving real processes by extracting knowledge from event logs (van der Aalst, 2011). As a discipline, process mining sits between, on the one hand, machine learning and data mining, and on the other hand, process modeling and analysis (van der Aalst, 2011). Some of the answers which process mining can deliver are: (a) what really happened, (b) why did it happen, (c) what is likely to happen in the future, and (d) when and why do organizations and people deviate (van der Aalst, 2011).

There are different algorithms used in process mining, depending on the data available and the questions that need to be answered. Irrespective of this, process mining requires structured data, and specifically, event logs of business (or other) processes. The goal is to analyze event data from a process oriented perspective (van der Aalst, 2011). For a process mining algorithm to work, a few attributes must be available. These are “case ID,” “activity,” and “timestamp.” Other attributes in the dataset give additional information on the process and can be used as well in specific types of analysis in process mining, but they are not critical to the analysis.

There are three types of process mining: discovery, conformance checking, and enhancement (van der Aalst et al., 2012). The most often used type of process mining is discovery (van der Aalst et al., 2012). This technique converts an event log into a process model, without any a-priori information (van der Aalst, 2011). The discovered model can be in the form of a Petri net, BPMN, EPC, or UML activity diagram, but it can also be a social network model, depending on the perspective needed (van der Aalst et al., 2012). Conformance checking uses an event log and a model as inputs. It is used for finding discrepancies between the reality (event log) and the process model (van der Aalst, 2011; van der Aalst et al., 2012). The third type, enhancement, also uses an event log and a model as an input, but the information from the event log is used to improve the existing process model (van der Aalst et al., 2012). Finally, process mining may refer to different perspectives of the analyzed processes. These are explained below.

Control-flow perspective—ordering of activities. Here, the goal is to find a good characterization of all possible paths by deriving a process model that provides the best summary of the flow followed by most or all of the cases in the event log (ProM, 2017a). It can answer questions such as:

There are several options for analyzing the case-flow. Some of the algorithms that can be used are the Alpha algorithm, the Heuristic Miner, Fuzzy Miner, and Inductive Visual Miner (IVM) (van der Aalst, 2016). A short comparison of these algorithms is presented in Table 1.

The answer of which algorithm should be used in a specific case is not a straightforward one. Table 1 provides a starting guideline when deciding which algorithm to use, but there are other options, and best is to test various algorithms and inspect the results. The Alpha algorithm, which was the first process mining algorithm developed, is not recommended for analysis of a real-world event log data (Rozinat, 2010). The Heuristic Miner was developed following the Alpha Miner to address its deficiencies and is therefore also able to simplify the process model by abstracting exceptional behavior and noise—by leaving out edges, i.e., connections between certain events (Rozinat, 2010). This algorithm is able to detect short loops and skipping of activities. However, it still shows rather complex process models (Buijs, 2017). The Fuzzy Miner interactively simplifies the process model by hiding some activities and paths, if desired (Rozinat, 2010). We use the Inductive Miner in our analysis because it was developed to overcome the disadvantages of other algorithms, and it shows a sound process model in the most user-friendly manner (Leemans et al., 2014).

Organizational perspective—focusing on information about the resources, which can be people, departments, roles, etc., and how they relate to each other. This relationship can also be represented as a social network based on the activities of the resources and can be used to find interaction patterns or evaluate the role of individuals (RapidProM, 2017). Social network mining is the most useful technique in the case of the organizational perspective, since network science is an area that studies interactions and relations between individuals. To derive sociograms from event logs, there are a few categories of metrics that have been developed (see Table 2).

In this paper, we use two metrics for the analysis: the Handover of Work and the Working Together metrics, because the information is displayed in a similar way. Thus, we believe that by showing the results from these two metrics, the reader will understand what type of information can be extracted and how it is visualized.

The metrics based on (possible) causality consider how work moves among performers (van der Aalst et al., 2005). In a football game, it considers the flow of events between the players. For instance, there will be a Handover of Work between two players if there are two subsequent activities/events where the first is completed by player A and the second by player B. In addition to a direct succession, it is also possible to analyze “indirect succession using a “causality fall factor” β, i.e., if there are 3 activities in-between an activity completed by i and an activity completed by j, the causality fall factor is β³” (van der Aalst et al., 2005, p. 9). The subcontracting metric counts the number of times when player B executed an activity in between two activities done by player A. For instance, Player A -> Player B -> Player A. This could indicate that work was subcontracted from Player A to Player B.

Metrics based on joint cases ignore the causality and simply count how often individuals are performing activities within the same case, i.e., sequence of activities (van der Aalst and Song, 2004). Thus, the metric Working Together shows which players most often participate or “work together” in the same ball possession sequence. If two individuals often work together on cases, they are considered to have a stronger relation than individuals rarely working together (van der Aalst and Song, 2004; van der Aalst et al., 2005).

Case perspective—focusing on the properties of the cases. It can answer questions, such as (ProM, 2017a):

Some options to answer the above questions with process mining are the Trace Variants, Dotted Chart visualizations, and the Trace and Sequence Clustering.

The basic idea of Trace Clustering is to split the event log into homogeneous subsets and for each subset to create a process model (Song et al., 2009). What this technique does is basically identification and clustering of similar sequences. The similarity is calculated based on a distance metric, usually the Euclidean or Hamming distance, while the clustering can be performed by using different algorithms, like k-Means or SOM (Veiga, 2009). A list of the algorithms available for clustering in ProM is presented in Supplementary Table 1.

Trace Clustering works by creating a set of profiles, each measuring a number of features for each case from a specific perspective (Song et al., 2009). In a second step, the distance between each case is measured by a distance metric; in this case, the Euclidean distance is used as it is found to be the most reliable. Finally, in a third step, similar cases are put together by using a clustering algorithm. Clusters can be analyzed independently from one another which improves the quality of the results for flexible environments (Song et al., 2009). Considering that football consists of 11 players who do not act according to a specific pre-defined process but rather based on quite a few distinct factors from their surrounding environment, one could reasonably assume that football can be considered a flexible environment within the process mining analytics area. Therefore, it would be interesting to see if and how trace/sequence clustering could be helpful for football performance analysis.

Sequence Clustering is based on a similar idea as Trace Clustering. However, this type of clustering is performed directly on the input data, i.e., no features are extracted from the sequences (Veiga, 2009). The plugin in ProM 5.7 has been implemented by Veiga (2009) whose algorithm is based on first-order Markov chains in which case the current state depends only on the previous state (Ferreira et al., 2007). The probability that an observed sequence is assigned to a given cluster is the probability that the observed sequence was produced by the Markov chain associated with that cluster, or simply the assignment of sequences to clusters is based on the probability of each cluster producing the given sequence (Ferreira et al., 2007; Veiga, 2009). Thus, a given sequence will be assigned to the cluster that is able to produce it with higher probability (Veiga, 2009). Veiga also adds two additional dummy states in the Markov chain: an input and an output state. This is necessary in order to represent the probability of a given event being the first or the last event in a sequence, which could be useful to distinguish between some types of sequences (Veiga, 2009).

In this paper, we use the SOM clustering algorithm and Markov chain clustering. SOM is used because it is very efficient with respect to computation time and is also quite robust concerning the results, especially for situations, where the characteristics of the process underlying an event log are largely unknown (Günther, 2009). The Markov chain clustering is preferred because it also discovers clusters without the analyst having to predefine the number of clusters. A detailed explanation of the algorithms is beyond the scope of this paper. For an overview of the SOM algorithm, the reader is referred to Si et al. (2003), and for more details on Markov chains, refer to Chung (1967).

Time perspective—analyzing the timing and frequency of events. If timestamps are available, it is possible to detect bottlenecks, monitor the utilization of resources, or predict the remaining processing time of running cases (van der Aalst, 2016). On its own, this perspective will most likely not be too interesting in a football scenario. However, combined with other perspectives, it can give interesting insights.

Each of these perspectives gives a different view of the process analyzed. The control-flow perspective relates to the “How” question, the organizational perspective to the “Who” question, while the case perspective answers the “What” question (ProM, 2017b). For a proper business understanding, users typically have to extract several models that describe different perspectives in the process analyses (Ingvaldsen and Gulla, 2008).

As seen, process mining is not a reporting, but an analysis tool, which is able to model and analyze complex processes (Rozinat and Gunther, 2015). Even though it works with historical data, it does not mean that it is limited to offline analysis, as the results can be applied to running cases (van der Aalst et al., 2012). Not all process mining types and perspectives can be applied to a football game scenario. From the three types of process mining mentioned, the discovery type is certainly applicable in this case, as conformance checking and enhancement require a model, which the discovered model from the OPTA log can be compared to. In football, there is no “perfect” or pre-defined model of the game process. Therefore, process mining can help with modeling the real-world process of what actually happened during the game. Finally, it is possible to view the event logs of the matches from all four perspectives discussed above.

Each of the process mining analytic perspectives explained previously (case-flow, case, and organizational) is used in the analyses introduced here.

In a first step, the original OPTA data are pre-processed and converted into an event log data. This is followed by analysis of the resulting event logs, discovery of the process models, and interpretation. The potential of the applied analytics techniques to gain a tactical understanding in football is presented in the Discussion section.

One exemplary game is analyzed as the goal is to demonstrate the techniques. The game between England and Iceland is chosen because Iceland won, while England's team showed one of its worst performances in a tournament. Therefore, it is interesting to explore what process mining can reveal about the tactical player and team behaviors. A summary of the game is presented in Supplementary Table 2.

The tools used for the analyses are:

This paper presents an exploratory study to evaluate the potential and suitability of process mining for tactical football performance analysis. As seen, process mining is a collection of algorithms and analytics techniques which are widely used in other domains for analyzing all kinds of business processes. It has never been applied specifically to sports, however. Not all algorithms and visualization techniques of process mining are demonstrated in this paper and not all types of process mining can be used for performance analysis in football. The discovery type of process mining algorithms makes the most sense, as it can demonstrate the exact behavior of teams and players. The conformance checking type of process mining is, in our opinion, not useful in this scenario because one does not have the perfect process model according to which players need to behave during the game. Enhancement of the process model does not seem to be useful in this case either. However, the discovery algorithms and techniques prove to be very valuable for analyzing a football game from a process perspective. Table 5 presents a summary of the techniques and algorithms addressed in this paper.

The case-flow perspective with the various types of algorithms for discovering the process model has the lowest potential to improve decision making in football based on event data. We can only detect which event types occur most frequently and how events are connected. In addition, for some event types, this perspective is not useful. For instance, the event “foul” will appear in both teams, and it is not clear from the mined process model which team has made or suffered how many fouls. This is due to the way in which the sequences are extracted from the original dataset. As each event has different qualifiers, if all those are considered when creating the process model, there would be too many variants. As the idea of the model is to give a quick overview of what happened as well as some dependencies between the activities, such a level of detail is not focused on in this paper. Unfortunately, it is not possible to avoid this issue. However, this is the case only for a limited number of events. This type of visualization, as demonstrated by the Inductive Visual Miner, can usually be used to analyze the process from a time perspective, i.e., to check which activities last too long and discover bottlenecks. However, due to the nature of football, analyzing this process model from the time perspective would make less sense. We consider the time perspective as part of the other two perspectives (case and social perspectives) which are more useful for performance analysis in football.

The case perspective offers various useful visual analytic techniques and clustering algorithms which can give valuable insights into both the team and player behavior. For instance, once the process model has been generated with the Inductive Visual Miner, it is possible to drill down and filter out specific sequences, which are of interest for the decision maker. We filtered out the attacking sequences which gives answers to questions like:

By using this option, it is possible to not only visualize the sequences leading to shot on goal for England's team but also to find out which players mostly started or ended a sequence. These analyses can be very useful in assessing the value of a player in a game.

Furthermore, clustering algorithms like SOM and first order Markov chains give a quick insight into the behavior of a team. Such analyses can be used, for example, during the half-time break in order to make tactical readjustments for the second half. Finally, social network analysis can be used for player insights. In this case, all event types occurring between the players are considered in the analysis in order to discover cooperation patterns between them. The two metrics that are applied, Handover of Work and Working Together, prove to be valuable in revealing important information about separate players. For instance, the Working Together metric can reveal which two players often cooperate in a sequence of ball possession, which in turn helps to plan tactical adjustments accordingly, especially concerning the defense of one own team. The Handover of Work metric can show which player is overwhelmed by having more work delegated from the other players. This could indicate fatigue or for the opponent can mean that that player should be the focus of their own defense.

We make an initial subjective evaluation of the potential usefulness of each technique presented in this paper for decision making in football. We base our evaluation on our knowledge of other data mining and visualization methods used in football performance analysis. However, this should be studied further by, for instance, conducting a study with coaches and other football experts, to gain an unbiased view on how useful such techniques are for actual decision making in football. Based on our initial results presented in this paper, process mining can be successfully used in addition to the traditional notational analysis for performance evaluation in football. It is an extension to the traditional analytics techniques that mostly consider the frequencies of actions. One major advantage is the possibility to visually infer patterns of interactions between the players and dependencies between different event types (goal, miss, duel, etc.), which has not been possible to achieve with other methods as, for instance, T-pattern analysis.

Based on the presented results, process mining offers valuable techniques and algorithms, which give insights into players' and team's behavior. The results are usually quick and understandable. This type of analysis can be used for examining successful and unsuccessful sequence outcomes, establishing defensive strategies against specific players, and overall gaining tactical insights into team and player behaviors. It is more user friendly compared to methods like T-pattern analysis. The sequences of events are clearer. There are also various options for additional analyses of the sequences as well as filtering out and focusing on specific types of sequences, e.g., offensive or defensive, sequences ending in a specific event, or sequences in which a particular player is involved, sequences that last longest, etc.

Process mining offers even more possibilities for analyses of the action sequences. Therefore, future research can explore whether the conformance checking type of process mining would be helpful in a football scenario. For instance, it may be possible to use conformance checking techniques to simulate and test the outcomes of sequences by enhancing the event log with other events. Finally, the pitch zones based on the positional data could be integrated to check if they give even more detailed insights from the sequence data.

The datasets for this article are not publicly available because the data is proprietary data and subject to costs. Requests to access the datasets should be directed to the author.

The paper is based on Chapter 8 of the first author's dissertation, which is also publicly available online (please refer to Kröckel, 2019).

PK wrote the paper and conducted the data analysis. FB revised the paper in terms of structure, content, writing style and grammar, acquired the data used in the analysis, and drafted the data analytics concept. All authors contributed to the article and approved the submitted version.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.