SDM 2012 Eamonn Keogh How to do good research, and get it published in top venues Solving Problems Now we have a problem and data, all we need to do is to solve the problem. Techniques for solving problems depend on your skill set/background and the problem itself, however I will quickly suggest some simple general techniques. Before we see these techniques, let me suggest you avoid complex solutions. This is because complex solutions... are less likely to generalize to datasets. are much easer to overfit with. are harder to explain well.
are difficult to reproduce by others. are less likely to be cited. Unjustified Complexity I From a recent paper: This forecasting model integrates a case based reasoning (CBR) technique, a Fuzzy Decision Tree (FDT), and Genetic Algorithms (GA) to construct a decision-making system based on historical data and technical indexes. Even if you believe the results. Did the improvement come from the CBR, the FDT, the GA, or from the combination of two things, or the combination of all three? In total, there are more than 15 parameters How reproducible do you think this is? Unjustified Complexity II There may be problems that really require very complex solutions, but they seem rare. see [a]. Your paper is implicitly claiming this is the simplest way to get results this good. Make that claim explicit, and carefully justify the complexity of your approach. [a] R.C. Holte, Very simple classification rules perform well on most commonly used datasets, Machine Learning 11 (1) (1993). This paper shows that one-level decision trees do very well most of the time. J. Shieh and E. Keogh iSAX: Indexing and Mining Terabyte Sized Time Series. SIGKDD 2008. This paper shows that the simple Euclidean distance is competitive to much more complex distance measures, once the datasets are reasonably large.
Unjustified Complexity III Charles Elkan Paradoxically and wrongly, sometimes if the paper used an excessively complicated algorithm, it is more likely that it would be accepted If your idea is simple, dont try to hid that fact with unnecessary padding (although unfortunately, that does seem to work sometimes). Instead, sell the simplicity. it reinforces our claim that our methods are very simple to implement.. ..Before explaining our simple solution this problemwe can objectively discover the anomaly using the simple algorithm SIGKDD04 Simplicity is a strength, not a weakness, acknowledge it and claim it as an advantage. Solving Research Problems Problem Relaxation: Looking to other Fields for Solutions: Wedont donthave havetime timetotolook lookatatall
all We waysofofsolving solvingproblems, problems,sosolets letsjust just ways lookatattwo twoexamples examplesinindetail. detail. look If there is a problem you can't solve, then there is an easier problem you can solve: find it. Can you find a problem analogous to your problem and solve that? George Polya Can you vary or change your problem to create a new problem (or set of problems) whose solution(s) will help you solve your original problem? Can you find a subproblem or side problem whose solution will help you solve your problem? Can you find a problem related to yours that has been solved and use it to solve your problem? Can you decompose the problem and recombine its elements in some new manner? (Divide and conquer) Can you solve your problem by deriving a generalization from some examples? Can you find a problem more general than your problem?
Can you start with the goal and work backwards to something you already know? Can you draw a picture of the problem? Can you find a problem more specialized? Problem Relaxation: If you cannot solve the problem, make it easier and then try to solve the easy version. If you can solve the easier problem Publish it if it is worthy, then revisit the original problem to see if what you have learned helps. If you cannot solve the easier problemMake it even easier and try again. Example: Suppose you want to maintain the closest pair of realvalued points in a sliding window over a stream, in worst-case linear time and in constant space1. Suppose you find you cannot make progress on this Could you solve it if you.. Relax to amortized instead of worst-case linear time. Assume the data is discrete, instead of real. Assume you have infinite space. Assume that there can never be ties. 1 I am not suggesting this is an meaningful problem to work on, it is just a teaching example Problem Relaxation: Concrete example, petroglyph mining I want to build a tool that can find and extract petroglyphs from an
image, quickly search for similar ones, do classification and clustering etc Bighorn Sheep Petroglyph Click here for pictures of similar petroglyphs. Click here for similar images within walking distance. The extraction and segmentation is really hard, for example the cracks in the rock are extracted as features. I need to be scale, offset, and rotation invariant, but rotation invariance is really hard to achieve in this domain. What should I do? (continued next slide) Problem Relaxation: Concrete example, petroglyph mining Let us relax the difficult segmentation and
extraction problem, after all, there are thousands of segmented petroglyphs online in old books Let us relax rotation invariance problem, after all, for some objects (people, animals) the orientation is usually fixed. Given the relaxed version of the problem, can we make progress? Yes! Is it worth publishing? Yes! Note that I am not saying we should give up now. We should still tried to solve the harder problem. What we have learned solving the easier version might help when we revisit it. In the meantime, we have a paper and a little more confidence. Note that we must acknowledge the assumptions/limitations in the paper SIGKDD 2009 Looking to other Fields for Solutions: Concrete example, Finding Repeated Patterns in Time Series In 2002 I became interested in the idea of finding repeated patterns in time series, which is a computationally demanding problem. After making no progress on the problem, I started to look to other fields, in particular computational biology, which has a similar problem of DNA motifs.. As happens Tompa & Buhler had just published a clever algorithm
for DNA motif finding. We adapted their idea for time series, and published in SIGKDD 2002 Tompa, M. & Buhler, J. (2001). Finding motifs using random projections. 5th Intl Conference on Computational Molecular Biology. pp 67-74. Looking to other Fields for Solutions You never can tell were good ideas will come from. The solution to a problem on anytime classification came from looking at bee foraging strategies. Bumblebees can choose wisely or rapidly, but not both at once.. Lars Chittka, Adrian G. Dyer, Fiola Bock, Anna Dornhaus, Nature Vol.424, 24 Jul 2003, p.388 We data miners can often be inspired by biologists, data compression experts, information retrieval experts, cartographers, biometricians, code breakers etc. Read widely, give talks about your problems (not solutions), collaborate, and ask for advice (on blogs, newsgroups etc) Eliminate Simple Ideas When trying to solve a problem, you should begin by eliminating simple ideas. There are two reasons why: It may be the case that that simple ideas really work very well, this happens much more often than you might think.
Your paper is making the implicit claim This is the simplest way to get results this good. You need to convince the reviewer that this is true, to do this, start by convincing yourself. Eliminate Simple Ideas: Case Study I (a) In 2009 I was approached by a group to work on the classification of crop types in Central Valley California using Landsat satellite imagery to support pesticide exposure assessment in disease. Vegetation greenness measure 190 180 Tomato Cotton 170 160 150 140 They came to me because they could not get DTW to work well..
130 120 110 100 0 5 10 15 20 25 At first glance this is a dream problem Important domain Different amounts of variability in each class I could see the need to invent a mechanism to allow Partial Rotation Invariant Dynamic Time Warping (I could almost smell the best paper award!) But there is a problem.
Eliminate Simple Ideas: Case Study I (b) It is possible to get perfect accuracy with a single line of matlab! Vegetation greenness measure 190 180 Tomato Cotton 170 160 In particular this line: sum(x) > 2700 150 140 130 Lesson Learned: Sometimes really simple ideas work very well. They might be more difficult or impossible to publish, but oh well. We should always be thinking in the back of our
minds, is there a simpler way to do this? When writing, we must convince the reviewer This is the simplest way to get results this good 120 110 100 0 5 >> sum(x) ans = 2845 10 15 20 2843 >> sum(x) > 2700 ans = 1 1 1 1
25 2734 1 0 2831 0 0 2875 0 0 2625 2642 2642
2490 2525 Eliminate Simple Ideas: Case Study II A paper sent to SIGMOD 4 or 5 years ago tackled the problem of Generating the Most Typical Time Series in a Large Collection. The paper used a complex method using wavelets, transition probabilities, multiresolution properties etc. The quality of the most typical time series was measured by comparing it to every time series in the collection, and the smaller the average distance to everything, the better. SIGMOD Submission paper algorithm Reviewers algorithm (a few hundred lines of code, learns model from data) X = DWT(A + somefun(B)) Typical_Time_Series = X + Z (does not look at the data, and takes exactly one line of code) Typical_Time_Series = zeros(64)
Under their metric of success, it is clear to the reviewer (without doing any experiments) that a constant line is the optimal answer for any dataset! We should always be thinking in the back of our minds, is there a simpler way to do this? When writing, we must convince the reviewer This is the simplest way to get results this good The Importance of being Cynical In 1515 Albrecht Drer drew a Rhino from a sketch and written description. The drawing is remarkably accurate, except that there is a spurious horn on the shoulder. This extra horn appears on every European reproduction of a Rhino for the next 300 years. Drer's Rhinoceros (1515) It Ain't Necessarily So Not every statement in the literature is true. Implications of this: Research opportunities exist, confirming or refuting known facts (or more likely, investigating under what conditions they are true) We must be careful not to assume that it is not worth trying X, since X is known not to work, or Y is known to be better than X In the next few slides we will see some examples If you would be a real seeker after truth, it is necessary that you doubt,
as far as possible, all things. In KDD 2000 I said Euclidean distance can be an extremely brittle distance measure Please note the can! This has been taken as gospel by many researchers However, Euclidean distance can be an extremely brittle.. Xiao et al. 04 it is an extremely brittle distance measureYu et al. 07 The Euclidean distance, yields a brittle metric.. Adams et al 04 to overcome the brittleness of the Euclidean distance measure Wu 04 Therefore, Euclidean distance is a brittle distance measure Santosh 07 that the Euclidean distance is a very brittle distance measure Tuzcu 04 Is this really true? Out-of-Sample 1NN Error Rate on 2-pat dataset Based on comparisons to 12 stateof-the-art measures on 40 different datasets, it is true on some small datasets, but there is no published evidence it is true on any large
dataset (Ding et al VLDB 08) True for some small datasets Almost certainly not true for any large dataset Euclidean DTW 0.5 0 0 1000 2000 3000 4000 Increasingly Large Training Sets
5000 6000 A SIGMOD Best Paper says.. Our empirical results indicate that Chebyshev approximation can deliver a 3- to 5-fold reduction on the dimensionality of the index space. For instance, it only takes 4 to 6 Chebyshev coefficients to deliver the same pruning power produced by 20 APCA coefficients Is this really true? 15 ty 10 5 32 16 0 64 8 128
256 Sequence Leng th 64 4 128 256 ali (Ding et al VLDB 08) 20 Di me ns i on No, actually Chebyshev approximation is slightly
worse that other techniques APCA light blue, CHEB Dark blue The good results were due to a coding bug.. .. Thus it is clear that the C+ + version contained a bug. We apologize for any inconvenience caused (note on authors page) This is a problem, because many researchers have assumed it is true, and used Chebyshev polynomials without even considering other techniques. For example.. (we use Chebyshev polynomial approximation) because it is very accurate, and incurs low storage, which has proven very useful for similarity search. Ni and Ravishankar 07 In most cases, do not assume the problem is solved, or that algorithm X is the best, just because someone claims this. A SIGKDD (r-up) Best Paper says.. (my paraphrasing) You can slide a window across a time series, place all exacted subsequences in a matrix, and then cluster them with K-means. The resulting cluster centers then represent the typical patterns in that time series. Is this really true? No, if you cluster the data as described above the output is independent of the input (random number generators are the only algorithms that are supposed to have this property ).
The first paper to point this out (Keogh et al 2003) met with tremendous resistance at first, but has been since confirmed in dozens of papers. This is a problem, dozens of people wrote papers on making it faster/better, without realizing it does not work at all! At least two groups published multiple papers on this: Exploiting efficient parallelism for mining rules in time series data. Sarker et al 05 Parallel Algorithms for Mining Association Rules in Time Series Data. Sarker et al 03 Mining Association Rules from Multi-stream Time Series Data on Multiprocessor Systems. Sarker et al 05 Efficient Parallelism for Mining Sequential Rules in Time Series. Sarker et al 06 Parallel Mining of Sequential Rules from Temporal Multi-Stream Time Series Data. Sarker et al 06 In most cases, do not assume the problem is solved, or that algorithm X is the best, just because someone claims this. Miscellaneous Examples Voodoo Correlations in Social Neuroscience. Vul, E, Harris, C, Winkielman, P & Pashler, H.. Perspectives on Psychological Science. Here social neuroscientists criticized for overstating links between brain activity and emotion. This is an wonderful paper. Why most Published Research Findings are False. J.P. Ioannidis. PLoS Med 2 (2005), p. e124. Publication Bias: The File-Drawer Problem in Scientific Inference. Scargle, J. D. (2000), Journal for Scientific Exploration 14 (1): 91106 Classifier Technology and the Illusion of Progress. Hand, D. J.Statistical Science 2006, Vol. 21, No. 1, 1-15 Everything you know about Dynamic Time Warping is Wrong.
Ratanamahatana, C. A. and Keogh. E. (2004). TDM 04 Magical thinking in data mining: lessons from CoIL challenge 2000 Charles Elkan How Many Scientists Fabricate and Falsify Research? A Systematic Review and Meta-Analysis of Survey Data. Fanelli D, 2009 PLoS ONE4(5) If a man will begin with certainties, he shall end in doubts; but if he will be content to begin with doubts he shall end in certainties. Sir Francis Bacon (1561 - 1626) Non-Existent Problems A final point before break. It is important that the problem you are working on is a real problem. It may be hard to believe, but many people attempt (and occasionally succeed) to publish papers on problems that dont exist! Lets us quickly spend 6 slides to see an example. Solving problems that dont exist I This picture shows the visual intuition of the Euclidean distance between two time series of the same length D(Q,C)
Suppose the time series are of different lengths? We can just make one shorter or the other one longer.. C_new = resample(C, length(Q), length(C)) Q C It takes one line of matlab code Solving problems that dont exist II But more than 2 dozen group have claimed that this is wrong for some reason, and written papers on how to compare two time series of different lengths (without simply making them the same length) (we need to be able) handle sequences of different lengths PODS 2005 (we need to be able to find) sequences with similar patterns to be found even when they are of different lengths Information Systems 2004 (our method) can be used to measure similarity between sequences of different lengths IDEAS2003 Solving problems that dont exist III
But an extensive literature search (by me), through more than 500 papers dating back to the 1960s failed to produce any theoretical or empirical results to suggest that simply making the sequences have the same length has any detrimental effect in classification, clustering, query by content or any other application. Let us test this! Solving problems that dont exist IIII For all publicly available time series datasets which have naturally different lengths, let us compare the 1-nearest neighbor classification rate in two ways: After simply re-normalizing lengths (one line of matlab, no parameters) Using the ideas introduced in these papers to to support different length comparisons (various complicated ideas, some parameters to tweak) We tested the four most referenced ideas, and only report the best of the four. Solving problems that dont exist V The FACE, LEAF, ASL and TRACE datasets are the only publicly available classification datasets that come in different lengths, lets try all of them
Dataset Resample to same length Trace Leaves ASL Face 0.00 4.01 14.3 2.68 Working with different lengths 0.00 4.07 14.3 2.68 A two-tailed t-test with 0.05 significance level for each dataset indicates that there is no statistically significant difference between the accuracy of the two sets of experiments. Solving problems that dont exist VI A least two dozen groups assumed that comparing different
length sequences was a non-trivial problem worthy of research and publication. But there was and still is to this day, zero evidence to support this! And there is strong evidence to suggest this is not true. There are two implications of this: Make sure the problem you are solving exists! Make sure you convince the reviewer it exists. Reproducibility Reproducibility is one of the main principles of the scientific method, and refers to the ability of a test or experiment to be accurately reproduced, or replicated, by someone else working independently. Reproducibility In a bake-off paper Veltkamp and Latecki attempted to reproduce the accuracy claims of 15 shape matching papers but discovered to their dismay that they could not match the claimed accuracy for any approach. A recent paper in VLDB showed a similar thing for time series distance measures. The vast body of results being generated by current computational science practice suffer a large and growing credibility gap: it is impossible
to believe most of the computational results shown in conferences and papers David Donoho Properties and Performance of Shape Similarity Measures. Remco C. Veltkamp and Longin Jan Latecki. IFCS 2006 Querying and Mining of Time Series Data: Experimental Comparison of Representations and Distance Measures. Ding, Trajcevski, Scheuermann, Wang & Keogh. VLDB 2008 Fifteen Years of Reproducible Research in Computational Harmonic Analysis- Donoho et al. Two Types of Non-Reproducibility Explicit: The authors dont give you the data, or they dont tell you the parameter settings. Implicit: The work is so complex that it would take you weeks to attempts to reproduce the results, or you are forced to buy expensive software/ hardware/ data to attempt reproduction. Or, the authors do give distribute data/code, but it is not annotated or is so complex as to be an unnecessary large burden to work with. We approximated approximated collections collections of of Explicit Non Reproducibility We time series,
series, using using algorithms algorithms time This paper appeared in ICDE02. The AgglomerativeHistogram and AgglomerativeHistogram and experiment is shown in its entirety, FixedWindowHistogramand and utilized utilized FixedWindowHistogram there are no extra figures or details. the techniques techniques of of Keogh Keogh et. et. al., al., in in the Whichcollections collections??How How the problem problem of
of querying querying Which the large?What Whatkind kindof ofdata? data? large? collections of of time time series series based based collections Howare arethe thequeries queriesselected? selected? How on similarity. Our results, indicate
on similarity. Our results, indicate that the the histogram histogram that Whatresults results?? What approximations resulting from approximations resulting from our algorithms algorithms are are far far superior superior our than those those resulting resulting from from the the than APCA algorithm algorithm of of Keogh
Keogh et. et. APCA al.,The superior superior quality quality of of our our al.,The superior by by how how superior histograms is is reflected reflected in in these these histograms much?, ,as as measured measured much? problems by by reducing reducing the the number
number problems how? how? of false false positives positives during during time time of series similarity similarity indexing, indexing, while while series How competitive?, competitive?, as as How remaining competitive competitive in in terms terms of of remaining measured how? how?
measured the time required to approximate We approximated approximated collections collections of of We time series, series, using using algorithms algorithms time AgglomerativeHistogramand and AgglomerativeHistogram FixedWindowHistogramand and utilized utilized FixedWindowHistogram the techniques techniques of of Keogh Keogh et. et. al., al., in in
the the problem problem of of querying querying the collections of of time time series series based based collections on similarity. similarity. Our Our results, results, indicate indicate on that the the histogram histogram that approximations resulting resulting from from approximations our algorithms
algorithms are are far far superior superior our than those those resulting resulting from from the the than APCA algorithm algorithm of of Keogh Keogh et. et. APCA al.,The superior superior quality quality of of our our al.,The histograms is is reflected reflected in in these
these histograms problems by by reducing reducing the the number number problems of false false positives positives during during time time of series similarity similarity indexing, indexing, while while series remaining competitive competitive in in terms terms of of remaining the time required to approximate
got aa collection collection of of II got opera arias arias as as sung sung by by opera Luciano Luciano Pavarotti, II compared compared Pavarotti, his recordings recordings to to my my his own renditions renditions of of the the own songs. songs.
My results, results, indicate indicate My that my my performances performances that are far far superior superior to to are those by by Pavarotti. Pavarotti. The The those superior quality quality of of my my superior performance is is performance reflected in
in my my reflected mastery of of the the mastery highest notes notes of of aa highest tenor's range, range, while while tenor's remaining competitive competitive remaining in terms terms of of the the time time in required to prepare Implicit Non Reproducibility
From a recent paper: This forecasting model integrates a case based reasoning (CBR) technique, a Fuzzy Decision Tree (FDT), and Genetic Algorithms (GA) to construct a decision-making system based on historical data and technical indexes. In order to begin reproduce this work, we have to implement a Case Based Reasoning System and a Fuzzy Decision Tree and a Genetic Algorithm. With rare exceptions, people dont spend a month reproducing someone else's results, so this is effectively non-reproducible. Note that it is not the extraordinary complexity of the work that makes this non-reproducible (although it does not help), if the authors had put free high quality code and data online Why Reproducibility? We could talk about reproducibility as the cornerstone of scientific method and an obligation to the community, to your funders etc. However this tutorial is about getting papers published. Having highly reproducible research will greatly help your chances of getting your paper accepted. Explicit efforts in reproducibility instill confidence in the reviewers that your work is correct. Explicit efforts in reproducibility will give the (true) appearance of value. As a bonus, reproducibility will increase your number of citations.
How to Ensure Reproducibility Explicitly state all parameters and settings in your paper. Build a webpage with annotated data and code and point to it (Use an anonymous hosting service if necessary for double blind reviewing) It is too easy to fool yourself into thinking your work is reproducible when it is not. Someone other than you should test the reproducibly of the paper. (from the paper) For double blind review conferences, you can create a Gmail account or Google Docs account, place all data there, and put the account info in the paper. How to Ensure Reproducibility In the next few slides I will quickly dismiss commonly heard objections to reproducible research (with thanks to David Donoho) I cant share my data for privacy reasons. Reproducibility takes too much time and effort. Strangers will use your code/data to compete with you. No one else does it. I wont get any credit for it. But I cant share my data for privacy reasons My first reaction when I see this is to think it may not be true. If you a going to claim this, prove it.
(Yes, prove it. Point to a webpage that shows the official policy of the funding agency, or university etc. Explain why your work falls under this policy) Can you also get a dataset that you can release? Can you make a dataset that you can publicly release, which is about the same size, cardinality, distribution as the private dataset, then test on both in you paper, and release the synthetic one? Reproducibility takes too much time and effort First of all, this has not been my personal experience. Reproducibility can save time. When your conference paper gets invited to a journal a year later, and you need to do more experiments, you will find it much easier to pick up were you left off. Forcing grad students/collaborators to do reproducible research makes them much easier to work with. Strangers will use your code/data to compete with you But competition means strangers will read your papers and try to learn from them and try to do even better. If you prefer obscurity, why are you publishing? Other people using your code/data is something that funding agencies and tenure committees love to see. Sometimes the competition is undone by their carelessness. Below (center) is a figure from a paper that uses my publicly available datasets. The alleged shapes in their paper are clearly not the real shapes (confusion of Cartesian and polar coordinates?). This is good example of the importance of the Send preview to the rival authors. This would have avoided
publishing such an embarrassing mistake. Alleged Arrowhead and Diatoms Actual Arrowhead Actual Diatoms No one else does it. I wont get any credit for it It is true that not everyone does it, but that just means that you have a way to stand above the competition. A review of my SIGKDD 2004 paper said (my paraphrasing, I have lost the original email). The results seem to good to be true, but I had my grad student download the code and data and check the results, it really does work as well as they claim. Parameters (are bad) The most common cause of Implicit Non Reproducibility is a algorithm with many parameters. Parameter-laden algorithms can seem (and often are) ad-hoc and brittle. Parameter-laden algorithms decrease reviewer confidence. For every parameter in your method, you must show, by logic, reason or experiment, that either
There is some way to set a good value for the parameter. The exact value of the parameter makes little difference. With four parameters I can fit an elephant, and with five I can make him wiggle his trunk John von Neumann Unjustified Choices (are bad) It is important to explain/justify every choice, even if it was an arbitrary choice. For example, this line frustrated me: Of the 300 users with enough number of sessions within the year, we randomly picked 100 users to study. Why 100? Would we have gotten similar results with 200? Bad: We used single linkage clustering...Why single linkage, why not group average or Wards? Good: We experimented with single/group/complete linkage, but found this choice made little difference, we therefore report only Better: We experimented with single/group/complete linkage, but found this choice little difference, we therefore report only single linkage in this paper, however the interested reader can view the tech report [a] to see all variants of clustering. Important Words/Phrases I Optimal: Does not mean very good We picked the optimal value for X... No! (unless you can prove it) We picked a value for X that produced the best..
Proved: Does not mean demonstrated With experiments we proved that our.. No! (experiments rarely prove things) With experiments we offer evidence that our.. Significant: There is a danger of confusing the informal statement and the statistical claim Our idea is significantly better than Smiths Our idea is statistically significantly better than Smiths, at a confidence level of Important Words/Phrases II Complexity: Has an overloaded meaning in computer science The X algorithms complexity means it is not a good solution (complex= intricate ) The X algorithms time complexity is O(n6) meaning it is not a good solution It is easy to see: First, this is a clich. Second, are you sure it is easy? It is easy to see that P = NP Actual: Almost always has no meaning in a sentence It is an actual B-tree -> It is a B-tree There are actually 5 ways to hash a string -> There are 5 ways to hash a string Theoretically: Almost always has no meaning in a sentence Theoretically we could have jam or jelly on our toast. etc : Only use it if the remaining items on the list are obvious.
We named the buckets for the 7 colors of the rainbow, red, orange, yellow etc. We measure performance factors such as stability, scalability, etc. No! Important Words/Phrases III Correlated: In informal speech it is a synonym for related Celsius and Fahrenheit are correlated. (clearly correct, perfect linear correlation) The tightness of lower bounds is correlated with pruning power. No! (Data) Mined Dont say We mined the data, if you can say We clustered the data.. or We classified the data etc Use all the Space Available Some reviewer is going to look at this empty space and say.. They could have had an additional experiment They could have had more discussion of related work They could have referenced more of my papers etc
The best way to write a great 9 page paper, is to write a good 12 or 13 page paper and carefully pare it down. You can use Color in the Text In the example to the right, color helps emphasize that the order in which bits are added/removed to a representation. In the example below, color links numbers in the text with numbers in a figure. Bear in mind that the reader may not see the color version, so you cannot rely on color. SIGKDD 2008 People have been using color this way for well over a 1,000 years SIGKDD 2009 Avoid Weak Language I Compare ..with a dynamic series, it might fail to give
accurate results. With.. ..with a dynamic series, it has been shown by  to give inaccurate results. (give a concrete reference) Or.. ..with a dynamic series, it will give inaccurate results, as we show in Section 7. (show me numbers) Avoid Weak Language II Compare In this paper, we attempt to approximate and index a d-dimensional spatio-temporal trajectory.. With In this paper, we approximate and index a ddimensional spatio-temporal trajectory.. Or In this paper, we show, for the first time, how to approximate and index a d-dimensional spatiotemporal trajectory.. Avoid Weak Language III The paper is aiming to detect and retrieve videos of the same scene Are you aiming at doing this, or have you done it? Why not say In this work, we introduce a novel algorithm to detect and retrieve videos.. The DTW algorithm tries to find the path, minimizing the cost.. The DTW does not try to do this, it does this. The DTW algorithm finds the path, minimizing the cost..
Monitoring aggregate queries in real-time over distributed streaming environments appears to be a great challenge. Appears to be, or is? Why not say Monitoring aggregate queries in real-time over distributed streaming environments is known to be a great challenge [1,2]. Avoid Overstating Dont say: We have shown our algorithm is better than a decision tree. If you really mean We have shown our algorithm can be better than decision trees, when the data is correlated. Or.. On the Iris and Stock dataset, we have shown that our algorithm is more accurate, in future work we plan to discover the conditions under which our... Use the Active Voice canbe beseen seenthat that ItItcan
Wecan cansee seethat that We seenby bywhom? whom? seen Experimentswere wereconducted conducted We Weconducted conductedexperiments... experiments... Experiments Takeresponsibility responsibility Take Thedata datawas wascollected collectedby
byus. us. The Wecollected collectedthe thedata. data. We Activevoice voiceisisoften oftenshorter shorter Active The active voice is usually more direct and vigorous than the passive William Strunk, Jr Avoid Implicit Pointers Consider the following sentence: We used DFT. It has circular convolution property but not the unique eigenvectors property. This allows us to What does the This refer to? The use of DFT? The convolution property?
The unique eigenvectors property? Check every occurrence of the words it, this, these etc. Are they used in an unambiguous way? Avoid nonreferential use of "this", "that", "these", "it", and so on. Jeffrey D. Ullman Motivating your Work If there is a different way to solve your problem, and you do not address this, your reviewers might think you are hiding something You should very explicitly say why the other ideas will not work. Even if it is obvious to you, it might not be obvious to the reviewer. Another way to handle this might be to simply code up the other way and compare to it. A Common Logic Error in Evaluating Algorithms: Part I
Here the authors test the rival algorithm, DTW, which has no parameters, and achieved an error rate of 0.127. They then test 64 variations of their own approach, and since there exists at least one combination that is lower than 0.127, they claim that their algorithm performs better Note that in this case the error is explicit, because the authors published the table. However in many case the authors just publish the result we got 0.100, and it is less clear that the problem exists. Comparing the error rates of DTW (0.127) and those of Table 3, we observe that XXX performs better Table 3: Error rates using XXX on time series histograms with equal bin size A Common Logic Error in Evaluating Algorithms: Part II To see why this is a flaw, consider this: We want to find the fastest 100m runner, between India and China.
India does a set of trails, finds its best man, Anil, and Anil turns up expecting a race. China ask Anil to run by himself. Although mystified, he obliging does so, and clocks 9.75 seconds. China then tells all 1.4 billion Chinese people to run 100m. The best of all 1.4 billion runs was Jin, who clocked 9.70 seconds. China declares itself the winner! Is this fair? Of course not, but this is exactly what the previous slide does. Keep in mind that you should never look at the test set. This may sound obvious, but I cannot longer count the number of papers that I had to reject because of this. Johannes Fuernkranz 0.8933 0.9733 0.9867 0.9333 0.9200 0.9200 0.9600 0.9600 0.9467 0.9200 0.9067 0.9067
0.9600 0.9600 0.9200 0.9200 0.9600 0.9467 0.9467 0.8933 0.9200 0.9200 0.9467 0.9200 0.9333 0.9333 0.9867 0.9200 0.9733 0.9333 0.9067 0.9467 0.9333 0.9467 0.9333 0.9600 0.9733 0.9333 0.9600
0.9467 0.9600 0.9733 0.9467 0.9600 0.9467 0.9467 0.9600 0.9333 0.9467 0.9200 ALWAYS put some variance estimate on performance measures (do everything 10 times and give me the variance of whatever you are reporting) Claudia Perlich SupposeIIwant wantto toknow knowififEuclidean Euclideandistance distanceor orL1 L1distance
distanceisis Suppose beston onthe theCBF CBFproblem problem(with (with150 150objects), objects),using using1NN 1NN best Better: Do 50 tests, report mean and variance 1 1 0.98 0.98 0.98
0.96 0.96 0.96 0.94 Accuracy 1 0.94 1 Red bar at plus/minus one STD 0.98 0.96 0.94 0.94
0.92 0.92 0.92 0.92 0.9 0.9 0.9 0.9 Euclidean L1 Euclidean L1 Euclidean L1
0.9733 0.9467 0.9600 0.9467 1.0000 0.9467 0.9733 0.9600 0.9600 0.9733 0.9867 0.9733 0.9333 0.9333 0.9600 0.9733 0.9600 0.9600 0.9733 0.9200 0.9333 0.9600 0.9733 0.9867 0.9867 0.9733
0.9733 0.9733 0.9333 0.9600 0.9200 0.9467 0.9333 0.9867 0.9867 0.9467 0.9867 0.9600 0.9867 0.9733 0.9867 0.9600 0.9467 0.9600 0.9733 0.9733 0.9733 0.9600 Much Better: Do 50 tests, report confidence
Accuracy A littler better: Do 50 tests, and report mean Accuracy Accuracy Bad: Do one test 0.9600 0.9600 Euclidean L1 229.00 166.26 170.31 163.61 179.06 170.52 164.91 168.69
164.99 184.31 189.76 170.95 168.47 164.25 178.09 178.53 166.31 Mean 175.74 STD 16.15 Variance Estimate on Performance Measures 230 230 220 220 210 210
Height in CM Height in CM Suppose I want to know if American males are taller than Chinese males. I randomly sample 16 of each, although it happens that I get Yao Ming in the sample Plotting just the mean heights is very deceptive here. 200 190 200 190 180 180 170 170 160
160 China US China US 167.08 166.60 161.40 175.32 173.31 180.39 182.37 177.39 167.75 179.81 174.83 171.04 177.40 166.41 180.62 173.00
6.45 Top Ten Avoidable s r e p a P s n o s a e R get Rejected, lutions To catch a thief, you must think like a thief Old French Proverb To convince a reviewer, you must think like a reviewer Always write your paper imagining the most cynical reviewer looking over your shoulder*. This reviewer does not particularly like you, does not have a lot of time to spend on your paper, and does not think you are working in an interesting area. But he/she will listen to reason.
*See How NOT to review a paper: The tools and techniques of the adversarial reviewer by Graham Cormode This paper is out of scope for SDM In some cases, your paper may really be irretrievably out of scope, so send it elsewhere. Solution Did you read and reference SDM papers? Did you frame the problem as a SDM problem? Did you test on well known SDM datasets? Did you use the common SDM evaluation metrics? Did you use SDM formatting? (look and feel) Can you write an explicit section that says: At first blush this problem might seem like a signal processing problem, but note that.. The experiments are not reproducible This is becoming more and more common as a reason for rejection and some conferences now have official standards for reproducibility Solution Create a webpage with all the data/code and the paper itself. Do the following sanity check. Assume you lose all files. Using just the webpage, can you recreate all the experiments in your paper? (it is easy to fool yourself here, really really think about this, or have a grad student actually attempt it). Forcing yourself to do this will eliminate 99% of the problems this is too similar to your last paper
If you really are trying to double-dip then this is a justifiable reject. Solution Did you reference your previous work? Did you explicitly spend at least a paragraph explaining how you are extending that work (or, are different to that work). Are you reusing all your introduction text and figures etc. It might be worth the effort to redo them. If your last paper measured, say, accuracy on dataset X, and this paper is also about improving accuracy, did you compare to your last work on X? (note that this does not exclude you from additional datasets/rival methods, but if you dont compare to your previous work, you look like you are hiding something) You did not acknowledge this weakness This looks like you either dont know it is a weakness (you are an idiot) or you are pretending it is not a weakness (you are a liar). Solution Explicitly acknowledge the weaknesses, and explain why the work is still useful (and, if possible, how it might be fixed) While our algorithm only works for discrete data, as we noted in section 4, there are commercially important problems in the discrete domain. We further believe that we may be able to mitigate this weakness by considering You unfairly diminish others work
Compare: In her inspiring paper Smith shows.... We extend her foundation by mitigating the need for... Smiths idea is slow and clumsy.... we fixed it. Some reviewers noted that they would not explicitly tell the authors that they felt their papers was unfairly critical/dismissive (such subjective feedback takes time to write), but it would temper how they felt about the paper. Solution Send a preview to the rival authors: Dear Sue, we are trying to extend your idea and we wanted to make sure that we represented your work correctly and fairly, would you mind taking a look at this preview there is a easier way to solve this problem. you did not compare to the X algorithm Solution Include simple strawmen (while we do not expect the hamming distance to work well for the reasons we discussed, we include it for completeness) Write an explicit explanation as to why other methods wont work (see below). But dont just say Smith says the hamming distance is not good, so we didnt try it you do not reference this related work. this idea is already known, see Lee 1978 Solution
Do a detailed literature search. If the related literature is huge, write a longer tech report and say in your paper The related work in this area is vast, we refer the interested reader to our techreport for a more detailed survey Give a draft of your paper to mock-reviewers ahead of time. Even if you have accidentally rediscovered a known result, you might be able to fix this if you know ahead of time. For example In our paper we reintroduced an obscure result from cartography to data mining and show (In ten years I have rejected 4 papers that rediscovered the Douglas-Peuker algorithm.) you have too many parameters/magic numbers/arbitrary choices Solution For every parameter, either: Show how you can set the value (by theory or experiment) Show your idea is not sensitive to the exact values Explain every choice. If your choice was arbitrary, state that explicitly. We used single linkage in all our experiments, we also tried average, group and Wards linkage, but found it made almost no difference, so we omitted those results for brevity (but the results are archive in our tech report). If your choice was not arbitrary, justify it. We chose DCT instead of the more traditional DFT for three reasons, which are
Not an interesting or important problem. Why do we care? Solution Did you test on real data? Did you have a domain expert collaborator help with motivation? Did you explicitly state why this is an important problem? Can you estimate value? In this case switching from motif 8 to motif 5 gives us a nearly $40,000 in annual savings! Patnaiky et al. SIGKDD 2009 Note that estimated value does not have to be in dollars, it could be in crimes solved, lives saved etc The writing is generally careless. There are many typos, unclear figures This may seem unfair if your paper has a good idea, but reviewing carelessly written papers is frustrating. Many reviewers will assume that you put as much care into the experiments as you did with the presentation. Solution Finish writing well ahead of time, pay someone to check the writing. Use mock reviewers. Take pride in your work!
Summary Publishing in top tier venues can seem daunting, and can be frustrating But you can do it! Taking a systematic approach, and being selfcritical at every stage will help you chances greatly. Having an external critical eye (mock-reviewers) will also help you chances greatly. Th e End