Again, the additional match data in combination
with handball specific knowledge help to transform
the data into a canonical format. Team tactics in team
handball are used to generate situations in which an
attempt has a high scoring probability. If the opponent
team is playing with a goalkeeper, then the high
scoring probability will only be achieved, if the
attempting player is in the same half of the field as the
opponent’s goal. There are cases in which the
attempting player is not in the half of the opponent’s
goal, but these cases are irrelevant for team tactics
because in these cases the opponent’s goalkeeper is
usually not present (or just about to return to the goal)
and thus no explicit tactics are applied.
Based on this observation, we can state that in
case of the relevant cases in the context of this work,
the attempting player needs to have a x-coordinate
between 0 and 20, given that the opponent’s goal has
a x-coordinate of 20. If the x-coordinate of the
attempting player is above 20, then we assume that
we need to transform the coordinates of all players
and the ball. I.e., we need a point reflection of the
coordinates using the centre of the field.
As a result, all coordinates are transformed such
that there are only attempts against the goal “of the
right side of field” (from the point of view of the
timekeeper of a match). Thus, the previously
described challenges two to four of section 2.2 are
resolved.
2.5 Sort Order of Player Coordinates
The concept of representing ttms as vectors has been
introduced previously (Schwenkreis, 2018a). In this
previous work, it has been proposed to use
classification based on deep learning. To overcome
the problem of “non-deterministic” ttm vectors, all
permutations of players of the ttm representing vector
were used to train the deep learning model. Now, if
clustering is used rather than classification, it does not
make sense to generate all permutations because it
would significantly distort the clustering result. It is
rather necessary to generate an order of players that is
well-defined.
From the point of view of team tactics there is no
need for a sort order across teams. It is rather
sufficient to have a well-defined sort order for each
team. Furthermore, it is irrelevant which sort order is
chosen as long as the sorting results in the same
sequence of player coordinates for all ttms that are to
be compared. Furthermore, it is important to ensure
that the vector position of a specific player remains
the same across all tposs of a ttm.
To determine the vector position of a player in
tposs, a heuristic is used that is derived from the
handball method to number the players by their
assigned offense position on the field: The “left-
wing” player is numbered one, the “half-left” player
two and so on. Finally, the goalkeeper gets number 7
and the ball number 8.
The offense position of players is defined by the
line-up data which is part of the additional match data
mentioned in section 2.3. For example, the player
who has been assigned to the “left-wing” position in
the line-up is assigned the vector position one as his
or her “coordinate index” in a tpos.
Some special cases need to be considered with
this approach: There might be the case when two
players with the same “nominal” offense position are
on the field, which would result in the same
coordinate index and an empty pair of coordinates in
the tpos. In this case, the y-coordinate in the starting
position of the tpos is used to determine the
coordinate index. There are three groups that are
handled separately: The two players with positions on
the left side, the three players in the mid and the two
players with positions of the right side.
In case of the offensive team, the player with
the highest y-coordinate is treated as the player
with the position defined in the line-up record.
Then the next empty coordinate slot in the same
player group of the tpos with a higher index is
used for the second highest y-coordinate.
In case of the defending team, the player with
the lowest y-coordinate is treated as the player
with the position defined in the line-up record.
The next empty coordinate slot in the player
group of the tpos with a lower index is used for
the player with the second lowest y-coordinate.
Cases with more than two players with the
same assigned position in the line-up are not
covered at this point.
3 TEAM MOVEMENT
SMIILARITY
Like classification clustering belongs to the family of
segmentation methods. The basic difference between
the two approaches is that clustering needs an explicit
notion of similarity (or distance), while classification
derives this notion implicitly based on the attribute
values of records with the same class label. Since the
assignment of class labels is very costly in the given
application scenario, the use of a non-supervised
approach based on clustering is proposed. Thus, a