Mining Evolution Patterns from Complex Trajectory Structures—A Case Study of Mesoscale Eddies in the South China Sea
Round 1
Reviewer 1 Report
The paper is well structured and comprehensive. However, there are some issues, which should be addressed.
General issues:
- There should be a review on the used language. Especially in the middle and end part (starting from section 2), there are some mistakes, which should be corrected. Further, please do not use abbreviations like “can’t”, “doesn’t” etc. It makes the article appear less formal.
- At least in my printed version the equations and figure 1 seem to have a quite low resolution.
- Equation numbers are missing.
Details:
- Figure 3b): For my understanding the transformed sequences should be V(V)SVSMV and MV(V)WVMV, respectively. If I am wrong, please correct me.
- How are parallel but differently behaving branches in a complex trajectory are represented in symbolic sequence?
- What is the difference between Ptype (eq. ?) and Ptk in CTSi? Does the transformation only consist of a discarding of the information besides the structural information?
- Keeping the location information would enable an analysis of the patterns in terms of their spatial distribution. For this purpose, a transformation of the trajectories is required which also considers the locations.
- In the domain of sequential pattern mining the terms “maximum” and “closest” pattern/itemset are already defined. You could also use them.
- Why do you first apply a clustering to your datasets and then mine the patterns in those clusters? Wouldn’t it be more interesting to first mine the patterns and then evaluate their spatial distribution? In this way, you could identify dominant regions for a certain kind of pattern and then draw your conclusions based on the results. You can further do it for n-longest patterns. In table 1, only the longest patterns are shown. What about the patterns which are slightly shorter but also interesting and behave differently?
Author Response
Dear reviewer,
Thanks for providing valuable comments on our manuscript. We break down your comments and provide our responses (dark red) to each of your comments separately as shown below. Changes are shown in bright red in the manuscript.
(1) There should be a review on the used language. Especially in the middle and end part (starting from section 2), there are some mistakes, which should be corrected.
Response: We have edited the texts and corrected the mistakes.
(2) Further, please do not use abbreviations like “can’t”, “doesn’t” etc. It makes the article appear less formal.
Response: All abbreviations like “can’t,” “doesn’t” etc. have been replaced.
(3) At least in my printed version the equations and figure 1 seem to have a quite low resolution.
Response: A high-resolution Figure 1 has been prepared and inserted into the manuscript.
(4) Equation numbers are missing.
Response: Equations are now orderly numbered.
(5) Figure 3b): For my understanding the transformed sequences should be V(V)SVSMV and MV(V)WVMV, respectively. If I am wrong, please correct me.
Response: In this manuscript, we represented the sequences by structures, such as the merging first and then splitting structure w, splitting structure s, merging structure m, and linear structure v. The splitting/merging points are a part of the corresponding splitting/merging structures and are not separately recorded.
(6) How are parallel but differently behaving branches in a complex trajectory are represented in symbolic sequence?
Response: The “parallel but differently behaving branches” are prioritized in the representation with the highest priority to form the merging first and then splitting structure w, then the splitting structure s, the merging structure m, and lastly the linear structure v. We have added texts to more clearly present how the “parallel but differently behaving branches” are represented in a symbolic sequence. (Lines: 185-187)
(7) What is the difference between Ptype (eq. ?) and Ptk in CTSi? Does the transformation only consist of a discarding of the information besides the structural information?
Response: Ptype refers to the type of a point along a complex trajectory. It could be either a(an) starting, ending, splitting, merging, merge–split point, or a common point. Any non-starting, ending, splitting, merging, split–merge points are defined as common points in this study (Lines 198-201). Ptk, on the other hand, refers to the structures of which the points change their types (Lines 177-181).
The transformation does contain only structural information, but we extracted the periodic pattern from the clusters pattern. The location information and structural similarities in fact has been considered in the clustering process.
(8) Keeping the location information would enable an analysis of the patterns in terms of their spatial distribution. For this purpose, a transformation of the trajectories is required which also considers the locations.
Response: Integration of the location information would definitely allow us to better illustrate the spatiotemporal patterns. In this study, we extracted the periodic pattern from the clusters of the ocean eddies. The location information in fact has been considered in the clustering process. We do agree that the location information should be integrated during the trajectory transformation process. In fact, we have made enough progress in expanding our method to primarily integrate the location information, but the results aren’t ready for publication.
(9) In the domain of sequential pattern mining the terms “maximum” and “closest” pattern/itemset are already defined. You could also use them.
Response: We have eliminated the definitions and added the references.
(Lines 217, 220)
(10) Why do you first apply a clustering to your datasets and then mine the patterns in those clusters? Wouldn’t it be more interesting to first mine the patterns and then evaluate their spatial distribution? In this way, you could identify dominant regions for a certain kind of pattern and then draw your conclusions based on the results. You can further do it for n-longest patterns. In table 1, only the longest patterns are shown. What about the patterns which are slightly shorter but also interesting and behave differently?
Response: The trajectory clustering considers both the location and the structural similarities of the trajectories. Extracting the frequent periodic pattern from the clusters thus allows us to loosely consider the location information of the trajectory structures. This at least partially remedies a drawback of our PPSE as mentioned in our response to Comment 8. We are improving our method so that we can mine the patterns across the SCS then examine their spatial distribution. We are working on a publication to compare the results derived from these two different perspectives.
It is true that there are n-longest patterns in addition to the longest pattern. We believe the longest pattern would be the most representative and is the best to reflect the structure changes of the trajectories. We do agree that it would be interesting to examine the other n-longest patterns.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
This paper presents a pattern minimg for Complex Trajectories.
The analysed trajectories are considered as complex because they can split and merge.
The proposed algorithm relies on projections into sequences of symbols.
Once converted, most frequent patterns are extracted from these symbolic sequences.
Threshold are used to detect frequent patterns.
The main analysis focuses on "semantic" pattern extraction.
l55 Author states that PrefixScan algorithm has "very high computational efficiency". However, this efficiency is not really discussed or supported by a reference, you could add few lines about its "efficency".
l150 AVISO dataset was downloaded but its usefullness in the process is not clear.
l198 The transformation into CTSi using a sequence of Pt is not clear for reviewer. Pt seems to represent the structural changes (v m s w) but there is also the "data items in the sequence". If there is 2 parrallel branch of a trajectory following a s and then these two branches are doing at the same time one split for branch 1 and 1 v for branch 2 (s + v != w), how do you decide to order the sequence ? This could be either : ssv or svs ?
Few typos were detected with missing words, extra letters or spaces see lines: 59, 168, 214
Author Response
Dear reviewer,
Thanks for you constructive and valuable comments, we have revised our manuscript accordingly. Our responses are in dark red. Changes are shown in bright red in the manuscript.
(1) (Line 55) Author states that PrefixScan algorithm has "very high computational efficiency". However, this efficiency is not really discussed or supported by a reference, you could add few lines about its "efficency".
Response: Please refer to Lines 58-61 for the new sentences we have added into the manuscript.
(2) (Line 150) AVISO dataset was downloaded but its usefulness in the process is not clear.
Response: In the discussion section, we used the AVISO data to validate our results and we have also added texts to explain how the AVISO data were used to validate our results in Lines 364-366. We have added a brief introduction of the AVISO data in Lines 370-374.
(3) (Line 198) The transformation into CTSi using a sequence of Pt is not clear for reviewer. Pt seems to represent the structural changes (v m s w) but there is also the "data items in the sequence". If there is 2 parrallel branch of a trajectory following a s and then these two branches are doing at the same time one split for branch 1 and 1 v for branch 2 (s + v != w), how do you decide to order the sequence ? This could be either : ssv or svs ?
Response: We have deleted the sentence “data items in the sequence” to eliminate the confusion.
Pt represents a structural change, i.e., change between a linear structure v, splitting structure s, merging structure m, and merging then splitting structure w. (Lines 177-181)
The second portion of this comment is very similar to Comment 6 from Reviewer 1. To tackle this situation, we prioritized the structures in the representation. The highest priority was given to form the structure w, then the structure s, the structure m, and lastly the linear structure v. (Lines 185-187)
(4) Few typos were detected with missing words, extra letters or spaces see lines: 59, 168, 214
Response: Yes, we have corrected the typos.
Author Response File: Author Response.pdf
