Here are the set of reviews for my ICSME’15 main track paper! Unfortunately, this bad boy was initially rejected from SANER’15, but we made many changes to the paper and it got in at ICSME. You can find a link to the PDF, code, and everything else in my publications.

SANER Rejection

Reviewer 1

This paper proposes to apply topic-modelling based information retrieval techniques (i.e., LDA and LSI) for feature location from the incremental changesets of source code. As an online learning algorithm based on changesets is adopted, it is not necessary to do retraining and get the updated topic models frequently. The authors further conduct evaluation on 14 open source Java projects to show the feasibility and effectiveness of the changesets approach.

Overall, this paper presents an interesting idea of using changesets for better feature location. Although LDA and LSI have been widely investigated in feature location domain, it is innovative to use the changesets from the version control system (e.g., SVN or Git) for feature location. The approach of modelling changeset topics is originally from reference [7]. This paper’s contributions mainly lie in the application of the approach of modelling changeset to feature location problem. The evaluation also seems to be solid. The authors publish the experiment data for public review. In the evaluation, it is good to use Wilcoxon signed-rank test with Holm correction to determine the statistical significance of the difference between results from LDA and LSI. However, as the authors mention in evaluation, few of the evaluated systems presented a statistically significant value between Snapshot based approach and changeset based approach.

The following issues need to be clarified: First, in this paper the authors use commit message in Git and SVN as the representation of a changeset in a version control system. Although the information among multiple versions of the project is used, the paper still focuses on feature location in a single version of the product. My concern is that if a feature that needs to be located is involved in several changes, how good can the proposed approach handle it? The authors may also better to show the effectiveness of the approach for features that are relevant and irrelevant to commit messages.

Second, the authors may also need to relate their work to the work on feature location on multiple versions of products. They may refer to the following literature and discuss about the application of modelling changeset topics for feature location in multiple versions.

Yinxing Xue, Zhenchang Xing, Stan Jarzabek: Feature Location in a Collection of Product Variants. WCRE 2012: 145-154

Third, for evaluation, the authors may try different parameter setup and measures of retrieval accuracy. Currently, the number of topic is set to 500. Actually, 500 topics may work well for normal documents based on natural language like English (see S.T. Dumais. LSI meets TREC: A status report, in Proceeding of Text Retrieval Conference, pp. 137-152. 1992), but a larger size of topic may be preferred for information retrieval on source code, considering more identifiers in source code. With regard to the measures, the authors only use the mean reciprocal rank (MRR). The authors may also consider some measures used in information retrieval domain, like Percentage of Relevant Queries (PRQ), Mean Average Precision (MAP) and Average Percentage of Code Units Investigated (APCUI). The different measures may reveal the different aspects of the results.

Below are also some detailed comments on the presentation and language of the paper:

  1. In introduction, in the last third paragraph, “Our results show that not only is our changeset approach feasible and practical, but in some cases out-performs current snapshot approaches.” Here, the authors should be more specific about the cases in which the proposed approach performs better.

  2. The approach section is a bit too simple. You may add more details, or merge it with some content in Section II.A and Section II.B.

  3. In the fourth paragraph of section IV.C, in the first sentence, “we our partitioning is inclusive of that commit.” should be “our partitioning is inclusive of that commit.”

Reviewer 2

Note: I had to leave this beauty verbatim…

Review follows A.J. Smith 4/90 IEEE/Computer 


Summary and Significance. 
 What is the purpose?   Is the the problem clearly stated? 
	Incremental modeling of text-based retrieval systems for program comprehension. This is a significant goal for the SANER audience.

Is there an early description of the accomplishments? 
          No. in particular, the authors fail to mention that the method works only sometimes. 

 Is the problem new?   Using I/R for program comprehension is not; the incremental change set approach is.

     Has the design been built before?  no

     Has the problem been solved before?  no

     Is this a trivial variation on or an extension of a previous result?  no

    Is the author aware of of related work? yes

     Does the author cite previous work and make distinctions from  it?  yes

    If an implementation, are there new ideas?  yes

 Is the method of approach valid?  yes

        Is the approach sufficient for the purpose? yes

        Sufficient discussion of new ideas? no; reasons for failure to reject null hypothesis need to be clarified.

Is the actual execution of the research correct? yes

      Algorithms correct? Convincing?  yes

      Did the author do what was claimed?  no

 Are the correct conclusions drawn from the results?  no

      What are the applications/implications of the results?  I'm not sure.

      Adequate discussion of these results? 
         There is discussion of what happened; not why it happened.

 Is the presentation satisfactory? 

      Readability? yes

      Does abstract describe the paper?  please use a structured abstract.

     Does the introduction describe the problem and the framework?  yes

     Appropriate amount of detail?  yes

     Figures/tables appropriate? too many.

     Self-contained? yes

Reviewer 3

The authors present an incremental topic-model approach to feature location based on change sets. They evaluate the technique by comparing changes sets, snapshots, and temporal change sets.

Excellent job sharing materials and making the work replicable by others.

Although the writing was clear, it was difficult to follow the thread of the research and how the study design answered the research questions. Especially missing are the big take away messages — what should a researcher or practitioner take away from this study in using change sets or snapshots for FLT?

Specific comments:

The explanation in section III seems unclear. Intuitively, I would think the topic model is run once on a snapshot, and then run incrementally on all the change sets after that point (up to the commit being searched). This approach is hinted at in the introduction (“online topic models can be instantiated once and incrementally updated over time.“) However, the wording in the following sentence:

“The changeset topic modeling approach requires two types of document extraction: one for the snapshot of the state of source code at a commit of interest, such as a tagged release, and one for the every changeset in the source code history leading up to that commit.”

Sounds like topic modeling is run on all the changes leading up to the snapshot. Is this the target usage scenario? Please clarify the writing to make the target usage scenario & algorithmic steps more clear. Figure 1 is a good start, but doesn’t clearly show how the change sets are involved. Figure 1 seems to show that the topic modeler is run on the whole snapshot every time, which I thought the purpose of the work was to avoid this?

I think the key insight behind the approach — “The key intuition to our approach is that a topic model such as LDA or LSI can infer any given document’s topic proportions regardless of the documents used to train the model.” — needs to be expanded. Isn’t this idea one of the main contributions of the work? A concrete example showing why this intuition is valid would help.

In section IV.C., the purpose of \theta_Queries is not yet clear, and it is difficult to see how this fits in to the larger study. It would be helpful if there were a big picture paragraph in the methodology section describing the parts of the study and how they are used to answer the research questions before diving in to the details. For instance, in this section I don’t yet know what temporal simulations look like, although that is one of the contributions of the work. It seems as if someone within the research team would perfectly comprehend section IV.C, but is not written so that a reader familiar with feature location can discern what is being evaluated and why when reading the paper from beginning to end.

Section IV.E: “To answer RQ1, we run the experiment on the snapshot and changeset datasets as outlined in Section IV-C. We then calculate the MRR between the two.” What two? How does this comparison help us answer RQ1? And then: “To answer RQ2, we run the experiment temporally as outlined in Section IV-C” the high-level goals of the temporal experiment and how it differs from a traditional experiment have not yet been described. Why are traceability links important to answering the research questions? It seems that the authors had some trouble making use of the Moreno data set. What is the advantage to keeping it in? More replications? Why include both Tables I & II, if only the data from Table II is used in the study?

It seems as if some of the high-level information I’m looking for might be partially buried in the discussion section in G, rather than being up front to help the reader understand the design of the experiment.

The work of Rao, et al. Seems closely related. In section II, can you differentiate why such an approach is less desirable than your proposed approach? (or evaluate?)


  • p. 5 C: “the process is varies slightly”
  • p. 5 C: “we our partitioning is“
  • conclusion: In this paper WE? conducted a study

(the mytical) Reviewer 4

Dear authors,

We would like to thank you for your submission that has lead to a lively discussion in the program committee. The main concerns raised by the committee pertain to:

  • the paper’s claim that the proposed approach analyzes multi-versions of changeset data, yet it seems that the paper did not really make good use of multi-version changeset data in the proposed approach and in the evaluation.

  • the fact that one of the reviewers familiar with this domain was not able to understand the approach since the paper has multiple issues making key points clear

ICSME Acceptance

Reviewer 1

The authors present a new approach in the context of feature location. They use information available in the a software configuration management system to incrementally perform concept location, so reducing time to perform such a kind of task. I found the idea behind the authors’ proposal very interesting even if it is not completely new in the context of software maintenance. The results support the validity of the new approach. The paper flow is adequate even if in some points I had some difficulties. For such difficulty I was not able to be completely confident with the the work done. Also, further details and justifications could be provided by the authors in the experimental part of the paper. All in all, I’m happy enough with the work done. It is one of the best papers I reviewed till now this year at ICSME.

In the following I’ll elaborate on the weakness points I see. I hope the authors will found them useful.

In the motivation part of the introduction, there are some points that seem contrasting each other. In particular, the authors wrote: “Indeed, given the current state-of-the-art in TR, it is impossible for an FLT to satisfy all three criteria while following the standard methodology.” Did Rao [10] and Hoffman at al. [9] make a contribution to satisfy all the three criteria? Reading the paper (and the Introduction, in particular) it seems YES.

Online (using fold-in and fold-out) LSI has been also applied in the context of architecture recovery. Mentioning this paper in the introduction section could further motivate your wok:

Michele Risi, Giuseppe Scanniello, Genoveffa Tortora: Using fold-in and fold-out in the architecture recovery of software systems. Formal Asp. Comput. 24(3): 307-330 (2012)

The part where the approach is highlighted in the introduction section needs to be rewritten because in the current form is not easy to follow. I read that paragraph more and more, but my comprehension level did not change: completely unclear.

Please discuss better on [10] and [18] in the related work section. In addition, it is not completely clear to me what the difference is between the proposed approach and [28].

Regarding the experimental part of the paper, I found very hard to understand the methodology (especially second paragraph). Last paragraph, the authors mentioned the dataset by Dit et al. Was the dataset by Moreno et al treated differently? Why?

Reading the description of the experiment, I was not able to understand whether the authors simulated the use of GitHub. I mean, were all the applications and the change sets in the used datasets in GitHub?

Last paragraph in section IV.E is not clear. I mean the place where the authors justify why RQ2 has been studied only on one dataset.

In section IV.F, the authors discussed on the fact that the p-value was greater than 0.05. In particular, they wrote: “This suggests that changeset topics are just as accurate as snapshot topics at the method-level, especially since there is a lack of statistical significance for any of the cases.” Since the null hypothesis has not been rejected, the authors can only discuss on descriptive statistics. That it, it seems that the authors accept the null hypothesis and this is definitively incorrect.

A statistical test (i.e., that chosen) verifies the presence of significant difference between two groups (in your case), but it does not provide any information about the magnitude of such a difference (if present). The magnitude of such a difference could be computed using a (non-parametric) effect size measure (e.g., Cliff’s d). You could also use the average percentage improvement/reduction.

Why the authors did not analyze execution time?

In the threats to validity you should also consider biases related to the statistical analysis performed (Conclusion validity). The readability could improve organizing threats in: Internal, External, Conclusion, and Construct.

Typing and formatting minor issues:

At the end of section III.C, there is (between brackets) a strange symbol.

Figure 2 is not compressible if the paper is printed black and white.

Please remove orphans.

Section 4.B - it is not so good reading the description of the experimental objects as the authors did.

Reviewer 2

This paper proposes a topic-modeling-based feature location technique in which the text retrieval model (i.e., topic model) is built incrementally from source code history. The technique uses an online learning algorithm to train topic models based on change sets, and thus can maintain an up-to-date model without incurring computational cost associated with retraining traditional snapshot-based topic models. The proposed technique has been evaluated and the results indicate that the accuracy of the technique is similar to that of a snapshot-based feature location technique.

This paper reports an interesting exploration of applying incrementally built topic models for feature location. It has the potential of improving current IR-based feture location methods with lower computational cost on building text retrieval models. But I think the paper still has a large space to improve.

First, the motivation of the paper is not clear and it is not well reflected in the evaluation. It seems that the main benefit of the proposed technique is the saving of computational cost associated with retraining traditional snapshot-based topic models. However, there is no analysis about how much computational cost can be saved. If the training of a snapshot-based topic model only takes a short time (e.g., several minutes), it is acceptable that the topic model is retrained for each release. Moreover, the saving of computational cost is not evaluated in the experimental study.

Second, the proposed technique is not well described. In the section presenting the technique (Section III), Section III-A and III-C respectively presents terminology and explains the reason why change set is used. Section III-B introduces the proposed technique, which is very short. Some important details are missing, for example how change set corpus are combined with snapshot corpus in training topic models? The process described in Figure 1 (B) does not reflect the incremental manner of the proposed technique.

Reviewer 3

The paper presents a topic-modeling-based Feature Location Technique (FLT) where, to reduce the computational cost, the model is updated incrementally from the changesets of commits from the project history instead of entire snapshots. The approach is evaluated on 1,200 defects on publicly available dataset (from 14 open-source Java projects) and is shown to exhibit accuracy not lower than the accuracy of more traditional models built on entire snapshots. The data and source code for the analysis are provided in an online appendix.

The idea is novel and the approach has potential. Not much work has addressed the issue of incremental model building in IR based feature location (the paper misses some related work – see below). The motivation behind building a model incrementally is to reduce the computational cost of rebuilding a model from every snapshot. The approach presented in the paper is sensible and the results indicate that it is a direction worth following. However, the paper also has several points where it needs some improvement.

The original motivation suggests that the changesets will update the model incrementally. My expectation was that every changeset will be considered separately, i.e., the model will be updated using a changeset. However, neither of the two research questions actually evaluates the approach in that setting. In RQ1 the changeset-based model is built using all changesets at once. In RQ2, the changesets are grouped into partitions based on the bug report that they are linked to and the model is updated using a partition. The first question here is why grouping changesets and why not updating the model after each commit? And then if a grouping is to be made, why not approximate a more realistic setting, i.e., update the model with every consecutive 10 commits, for example. Consecutive commits will address different bugs and thus will certainly have different topic distribution. My doubt here is to what extent the grouping in RQ2 may have introduced a bias in the results? By the w! ay, the part describing the historical simulation is somewhat confusing – at least I had to read it twice to fully understand what exactly is being done.

When investigating the accuracy of the models built on the changesets the thresholds are selected without justification and no tuning. For instance, for the number topic models in all analyzed projects is fixed to 500. The paper justifies the lack of parameter tuning with the fact that the “goal is to show the performance of the changeset-based FLT against snapshot-based FLT under the same conditions” and that “the measurements collected are fair and that the results are not influenced by selective parameter tweaking”. However, poor selection of the parameters may lead to poor results and thus unrealistic optimism that the proposed changeset-based FLT performs as good as traditional snapshot-based FLTs. This doubt is somewhat confirmed by results shown in Tables 1 and 2: The Mean Reciprocal Rank (MRR) is used to measure the effectiveness of a FLTs for a set of queries; the higher the value the better the result. The values for MRR shown in Tables 1 and 2 are quite low and this is true for both models. For example, for the project Pig v0.8.0, the MRR is ~ 0.011. This score of MRR would mean that the minimum rank for a relevant class would be on average ~ 90 (out of 442 classes in this project). The MRR reported by Moreno et al. varies depending on the settings and type of information that is considered but stays between 0.18 and 0.26 for the same project. This corresponds to ranks 6 and 4 (again out of 442). Thus, the doubt here is that the results of the snapshot-based FLT using the selected parameters are poor and the only thing that one can conclude is that the changeset-based FLT is not making the poor results worse. Now whether the poor results are due to the underlying techniques (i.e., LDA and-or LSI) or to the parameter selection only is not clear but is probably worth investigating.

RQ1 should be rephrased maybe as a hypothesis “Changeset-based FLT is less accurate than snapshot based FLT”. Then the data shows that this cannot be proved.

Regarding RQ2, it is not clear how the accuracy’s “fluctuation” of the CFL technique is measured as a project evolves. I do not think the MRR metric by itself measures such fluctuation, or at least this is not explained in the paper. The MRR only measures accuracy. I would think that series analysis on the MRRs across time would be the way to go or other analysis of this kind. Now, it seems that the goal was onlu to compare the accuracy when changesets are used to incrementally update the topic model, as opposed to update the model at once with all the changesets. Unfortunately, it is not clear whether what the goal really is. I suggest to clarify this issue and perhaps reformulate RQ2. After all, the main goal of the paper is to test how the CFL would perform in a realistic environment where the model is incrementally updated with changes in commits.

The paper omits the LSI results “for brevity”. If they are omitted completely, it is best not to even mention them. The best thing to do is to at least mention how they compare wrt LDA.

Detailed comments:


  • “By training an online learning algorithm using changesets, the FLT maintains an up-to-date model without incurring the non-trivial computational cost associated with retraining traditional FLTs.”: As shown in Fig. 1 the snapshots are still used for indexing. Thus, the computational cost is saved in the process of building the topic model. What is exactly the saved computational cost? To better motivate the paper I would recommend to give a citation or an example of how long it takes to create a topic model for a large system such as eclipse using LDA. Also, it is a good idea to provide the cost saving of the Online LDA technique, compared to the standard LDA.
  • “It follows from the first two observations (1: Like a class/method definition, a changeset has program text; 2: Unlike a class/method definition, a changeset is immutable.) that it is possible for an FLT following our methodology to satisfy all three of the criteria above. “: It is not clear how the first criterion is satisfied, i.e., “(1) accurate like a TM-based FLT”
  • “We then used a subset of over 600 defects and features to conduct a historical simulation that demonstrates how the FLTs perform as a project evolves.”: Why 600?
  • The preprocessing often includes stemming, but stemming is not mentioned here. Later (p.6, Section IV Study) it becomes clear that no stemming is applied without justifying why.


  • “Normalizing: replace each upper case letter with the corresponding lower case letter”: Lawrie et al. use “normalization” for vocabulary normalization (i.e., the alignment of the vocabulary found in source code with that found in other software artifacts). See: D. Lawrie, D. Binkley, and C. Morrell. Normalizing source code vocabulary. In Proceedings of the Working Conference on Reverse Engineering (WCRE), pages 3-12, 2010
  • “corpus is a set of documents (i.e., methods)”: “i.e.,” -> “e.g.,”


  • Section IV.C. (Methodology) can be broken down into subsections based on the RQs.
  • To answer RQ2 (Does the accuracy of a changeset-based FLT fluctuate as a project evolves?), the paper describes the so-called historical simulation where commits are related to each query (or issue) and partitions of mini-batches of changesets are created. The model is then updated using a mini-batch. An index of topic distributions with the snapshot corpus is then inferred. I don’t understand why for the historical simulation, commits are grouped into partitions of mini-batches instead of updating the model after every commit.
  • “on all documents extracted.” -> extracted documents


  • The paragraph starting with “After extracting tokens, we split … “ is not needed. The preprocessing, except the stemming, is already explained in Section II.A.
  • Thresholds are missing justifications: K, the number of topics, is chosen to be 500; the two parameters that control how much influence a new mini-batch has on the model when training are 1024 and 0.9. No justification is given for the selected values. What are the values selected in related works?


  • Ref. [2]: the publication date is 2013.
  • The references should be consistent. For example, the venue of the references 7, 17, 19 and 20 have the following form: “Software Engineering, IEEE Transactions on”; instead of “IEEE Transactions on Software Engineering”.

Missing references to related work:

Hsin-yi Jiang, Tien N. Nguyen, Carl K. Chang, and Fei Dong, “Traceability Link Evolution Management with Incremental Latent Semantic Indexing”, in Proceedings of the 31st IEEE International Computer Software and Applications Conference (IEEE COMPSAC 2007), pages 309-316, July 24-27,2007

Hsin-yi Jiang, Tien N. Nguyen, Ing-Xiang Chen, Hojun Jaygarl, Carl K. Chang, “Incremental Latent Semantic Indexing for Automatic Traceability Link Evolution Management”, in Proceedings of the 23rd ACM/IEEE International Conference on Automated Software Engineering (ACM/IEEE ASE 2008), September 15-19, 2008

Ratanotayanon, Sukanya, Hye Jung Choi, and Susan Elliott Sim. “Using transitive changesets to support feature location.” Proceedings of the IEEE/ACM International Conference on Automated Software Engineering, 2010