Pattern analysis

4.4.1 Methodology: Sequential association rule mining

Thus, this study adopted the technique of sequential association rule mining to mine the patterns of attention allocation composed of sets of jointly occurring renewal cycles. This method has long been effectively applied to examine factors contributing to crash occurrence (Geurts et al. 2005, Pande and Abdel-Aty 2009, Montella 2011, Montella et al. 2012). The association rule mingling is the technique of identifying the co-occurrence relation among several items. Set of item that usually appear jointly can be extracted. In addition to the co-occurrence relation, the sequential association rule mining includes the time dimension. That is, this method did not only explore the items that jointly appear, it also identifies the sequence of item appearing. Following sections will introduce a brief concept of the sequential rule mining. A detailed description of the method can be found in, Introduction to Data Mining, by Tan et al.

(2006).

Before the introduction of the method, terminologies used in this study and in sequential rule mining must be defined and differentiated. Table 4-5 shows the three terminologies used in the sequential rule mining and driver attention allocation model.

Table 4-5 Definition of sequence, element and event

Sequential Rule Mining Driver Attention Allocation Model Event (Item) The basic component of rule. Purpose of

this method is to identify the sequence of event appearing.

Focal points are the basic component for exploring pattern of attention allocation.

Element (Transaction)

A set of event appearing at specific time. The element refers to the glance in driver attention allocation. Since drivers can only glance at only one focal point, an element can only include one event.

Sequence An order list of elements. Is can be characterized by its size. A k-sequence indicated a sequence with k items.

A pattern of attention allocation, which is a sequence of focal points that drivers glance.

The sequential rules are generated by merging the existing sequence. The a priori principle stated that a k-sequence can be derived by combining two (k-1)-sequence. Assuming that there are (k-1)-sequence1 and (k-1)-sequence2 exist in the rule set. If the sub-sequence of dropping the last event of (k-1)-sequence1 is identical with another sub-sequence of dropping the first event of (k-1)-sequence2, these two (k-1)-sequences can be merged as a new k-sequence. Then, the newly generated sequences are included in the rule base, if they satisfy the minimum performance requirement. Otherwise, the sequences are pruned.

In the sequential association rule mining method, two performance measurements, support and confidence, are often used. Support value determines how often a combination of renewal cycles (a rule or pattern) can be found in the entire data set. As shown in Eq. (5), it is the percentage of events in entire data set covered by the rule of X → Y.

𝑠(X → Y) =^σ(X→Y)_N (6) 𝑠(X → Y) represents the support value of a renewal cycle X appearing before a renewal cycle Y; σ(X → Y) represents the number of events that fit the rule of X → Y ; and N is the total number of events. Higher support value suggested important patterns that appear more frequently and explain more about the data set.

As shown in Eq. (6), confidence value is the percentage of events having renewal cycle X that also contains a renewal cycle Y.

𝑐(X → Y) =^σ(X→Y)_σ(X) (7) 𝑐(X → Y) is the confidence of rule X → Y; σ(X → Y) is the number of events which fit the rules X → Y; σ(X) is the number of events containing a renewal cycle X in the data set. This research, the maximum gap between two renewal cycles was set at one. That is, the renewal cycle Y appears right after X.

The confidence value was used in this research to derive the sequential rules of renewal cycles. Higher confidence suggested the higher strength between the renewal cycles. However, some rules with high confidence may have only few samples and cannot efficiently explain the entire data set. Therefore, a minimum support should be set to filter the rules. Referring to previous research (Geurts et al. 2005, Pande and Abdel-Aty 2009, Montella 2011), this study set the minimum support at 5% and the minimum confidence at 10%.

4.4.2 Pattern generation

One might ask whether the generated renewal cycles were interrelated or not. To answer this question, the Sequential Association Rule Mining package in SAS Enterprise Miner 6.2 was used to mine the sequential association between renewal cycles and combine related cycles into patterns of attention allocation. As stated, drivers displayed different renewal cycles under various road conditions and with various driver intentions. Hence, driving straight on a segment, passing through an intersection, and changing lanes on a segment were separated to mine the respective sequential association rules. Other types of maneuvers were not included because of the small sample size. Table 4-6 shows the derived attention allocation rules of the sample drivers for the three maneuver intentions.

Table 4-6 Attention allocation patterns of various maneuver intentions

Patterns of

*The confidence and support values are expressed as percentages -No pattern found

The renewal cycles that included the ReM occurred in almost all extracted patterns of attention allocation. This finding suggests that paying attention to the front and rear areas of the vehicle were the two most crucial components for observing the surrounding traffic and maintaining situational awareness. The sample drivers usually transited their vision to these two areas immediately before or after shifting their attention elsewhere.

Driving straight on a road segment is relatively simple and has a light information load. In addition to the mentioned crucial renewal cycles, the drivers traveling straight on a road segment displayed the pattern related to InvD and LW glances. These cycles containing non-driving related information and, when combined with the above-mentioned crucial cycles, formed the attention allocation pattern for driving straight on a road segment. Such an attention allocation pattern can be described as drivers comfortably focusing on the front and rear areas, but intermittently and casually directing their attention to distractions on the roadside or inside the vehicles. The drivers also displayed the cautious behavior of transiting their vision from the LM to the ReM to maintain their situational awareness of the rear area, probably to monitor the blind zone on the left side.

For the maneuver intention of passing through an intersection, the drivers experienced a relatively heavy task load because of possible threats arising from the intersecting traffic. Compared with driving straight on a road segment, fewer notable patterns of renewal cycles were evident because the drivers were more cautious and concentrated on a few critical focal points when passing through an intersection.

Apart from concentrating on forward and backward areas, the renewal cycle pattern F-RF→F-LW showed that the drivers did not transit their vision far from one side of the vehicle to the other side, i.e., renewal cycle F-RF-LW. An intermediate glance at the forward side was adopted. The sample drivers usually looked to RF field initially, where conflicts with right-turning traffic would occur. Afterward they turned their vision to the LW to check for possible traffic emerging from the intersected road.

When changing lanes, drivers may encounter threats from multiple directions and must expend heightened effort to prevent possible conflicts, particularly conflicts from the adjacent lanes. Under these intense circumstances, the sample drivers’ InvD were minimized and attention to the rear and side areas was strengthened. This finding suggests that the ReM was used in an auxiliary manner to enhance the drivers’

situational awareness of the rear area, and that the LM was used to monitor the blind zone. Compared with the intentions of driving straight on a segment and passing through an intersection, the drivers evidently considered changing lanes to be a more mentally demanding task. Thus, after a renewal cycle for InvD, a relatively high proportion (42.86%) of the drivers immediately transited their vision to the ReM to gain information of the rear area relevant to changing lanes.