As almost always, the answer is "it depends" :-)

I think there are a few major factors on which the answer depends:

1. on the amount of data distortion (are most of the men or women lying about their age, and by how much) and

2. the particular DM application and its dependence on the shape and in particular the tails of the distribution of the data.

For example, in recommender system applications, the influence a few colluding users may have in distorting recommendations (e.g. by providing very low ratings on certain (new) items, in the hopes of promoting other items for which they have a vested interest) is next to null when the dataset which is used to build the models is really "Big". On the other hand, in other analytics applications, where an empirical distribution is used, and where it's the tail of the distribution that determines the results (think inventory control applications, or yield management in the airline sector, and other similar areas), then even a small amount of data distortion (e.g. artificial introduction of small amounts of unusually high demand), can significantly alter the results of the DM process.

On the other hand, some times, after some point, increased data distortion may lead to higher accuracy of the results: in your example, if practically everyone lies about their age, and they all do it more or less in the same way (say subtract more or less 10% off their real age), then a classifier will probably work as well as if nobody lied about it: the system will account for the "age deflation" phenomenon, and will provide pretty accurate results. The results will be much worse if only about half the population lies about their age...

Lying obviously affects the results of data mining. In some cases, it justifies an entire analytics industry. Usually it is dressed up in the word "fraud" in those cases, but the essence is the same.

In simpler situations, it is often useful to make it relatively easy to lie blatantly to make it easier to remove untruthful answers. Defaulting gender to none, country to Albania and age to 99 allows you to remove people who don't want to disclose demographics relatively easily.

There is also a thing called Bayesian truth serum. It is based on the fact that people over-estimate the prevalence of characteristics that they themselves have if that characteristic is not externally obvious. Thus, Catholics estimate more of the population is Catholic than is true and Jews over-estimate the proportion that is Jewish.

With Bayesian truth serum, you simply ask people what the percentage of the population is who would lie on the form in question. If they estimate a higher proportion, they are likely to be lying themselves. Obviously, this isn't guaranteed, but it can be surprisingly reliable.

Very interesting topic, and it seems that there's been a good deal of interest of late, especially in relation to social networks, to try and detect when people are lying/exaggerating about things.

http://www.theguardian.com/uk/interactive/2011/dec/07/london-riots-twitter

http://mashable.com/2014/02/25/pheme-twitter-lie-detector/

This EU-funded PHEME project looks to be tackling the particular issue of rumours and lies on social media.

http://www.pheme.eu/

Veracity appears to be one of the least consulted V's of the big data canon, at least up until recently.

Most people will cheat if they think that they can get away with it. The nice thing about data mining is that it gets the cheating patterns even better when there is an attempt to hide them -- by using the special traits of the disguise method :)

GIGO - applies everywhere, not just big-data.

Garbage In - Garbage Out.

We had first-hand experience of this with few of our customers.

To Ted's observation: detecting fraud is a common application of DM. However, there one does not seek to eliminate the "lying" prior to mining but to detect it through mining. It seems that the issue here is somewhat different ... The lying is confounding the rest of the analysis ... I.e negatively impacting the goal of the mining exercise. Unlike the previous case where we have known instances of fraudulent and not fraudulent instances, here the data set does not clearly label who is a fraud. As such, some amount of cleansing is required prior to mining. In that sense the two scenarios are different. One may make the case that one is mining to detect and eliminate fraudulent responses prior to mining for matchmaking, though, but the end goal is matchmaking. And, to the original question, I'm sorry I do not know of actual anecdotal instances of this phenomenon but imagine it must be very common where data is acquired through polling, questionnaires, interviews, etc. --- I.e. Where knowledge elucidation must rely on what an individual expresses orally, in writing and such. I am no expert but colleagues who conduct survey-based research tell me survey design takes into account the possibility that respondents could be "inconsistent" in their responses and tries to account for it. Not sure if matchmakers use such questionnaire design strategies, though.

I meant to say "knowledge elicitation" but my smart phone thinks it knows better!

I agree with Ram about the similarity of Lying to Fraud detection, and that the analyst needs to know in advance which are the wrong records, in order to train the data mining algorithm.

It seems that Lying and Fraud are actually the same, except maybe their complexity - business frauds are more complex.

BTW, Fraud data is not as complete as it may look like, many illegal activities go below the system's "radar". On the other end, Lying has indications (for example complaints and irregularities) which are as good as formal input.

Edith Ohri and Ram Pakath: I agree that fraud and spam is deliberate and intentional, just like deception. But deception does not necessarily carry the same connotation of securing unlawful gain. For instance, a colleague of mine tells me a quick anecdote about OKCupid. Apparently, according to her friend, there is a 'body type' category on OKCupid of 'used up'. She says they "were laughing about it ever since. A few years ago, after a divorce, [her friend] tried dating on OKCupid and chose that body type in her profile, more as a signifier of her sense of humor and value system than anything reality based." Did she lie? Well, she most certainly did. Did she defraud anyone? I doubt it. Deception has lots of permutations. Another example is lying in the context of concealing a brithday party. Quite benevolent isn't it? Deceivers' intentions are not always as easy to decode, while spammers, or say insurance fraudsters, are clearly after a certain kind of gain.

Ioannis Christou: I agree, the consistency of lying is important and is largely unknown. Data is "is not always ... X" is not the same as "is always ... not X". I wonder if DM is at all the right tool for figuring it out...

Ted Dunning: Bayesian truth serum idea is brilliant! I'll look around to see what, if anything, has been done in automated deception detection. I'd be very curious to see how reliable it is. Have you seen any articles testing it, by any chance?

Gopalakrishna Palem: I think the key in avoiding GIGO is figuring out that G. What sort of data are your customers providing you with? No specifics, just wanted to get an idea of the topic or type of business.

Gerard Lynch: thanks for the link to the PHEME Project. It's rather new (only 3 months old or so, it appears) and I'll keep track of their progress. Do you have any direct links to the group, by any chance?

Two part answer and an observation:

a) The key words in your question are "BIG DATA". Big data is not simply data mining, it is the data mining of _extremely_ large data sets. A survey of 10,000 is not big data, and a survey of 1 million is at the very low end of the big data scale. The key implicit benefit of big data is that the large numbers compensate for the data quality. One or ten or a hundred thousand lies is not a big factor on a scale of millions of surveys;

b) Without clear context (your prelude story is about small data while you ask about big data) it occurs to me that you could be indirectly asking about the effect of lies on small data. In that case, it all depends on the modelling and data cleaning approaches used. For example, support vector machines are very sensitive to bad data while random forests and ensemble approaches are much less so. Data cleaning (if done) is usually better, albeit slower, if there is a trained human-in-the-loop. Automated cleaning can introduce artifacts of its own.

Observation on match-making sites: If everyone is lying, that becomes the truth. Participants will read answers and mentally adjust (eg.- 'He's probably shorter than he says.'), consequently any modelling should do that as well. If the truth is relative, why fuss about absolutes? The point of the sites (I thought) is to find good matches, not the 'truth'.

Mike Bourassa: Well, I get the point with the big vs. small data, but I don't get how more observations would compensate for the quality and improve the match. The 3 predictive questions that couples agree on were claimed to be:

"Do you like horror movies?"

"Have you ever travelled (sic.) around another country alone?"

"Wouldn't it be fun to chuck it all and go live on a sailboat?"

These are yes/no questions, i.e., binary answers are expected and in fact reflect one's preferences. So, how do you suppose the participants should mentally adjust? So, if she says she loves horror movies, she really means she hates them?

What does "compensating for quality" actually mean?

Hi Victoria,

I don't have any direct links to the PHEME project that I am aware of, although it looks like a subject that I would definitely like to get involved in at some point, having carried out some Twitter analytics recently. I came to it via a BBC news article actually,

http://www.bbc.com/news/technology-26263510

in a manner similar to the initial question that you posed.

One area I found interesting previously was a study on pastiche detection from an authorship attribution point of view,

http://dl.acm.org/citation.cfm?id=2388627

I guess maybe I've drifted away from the initial point a little bit, but you have touched upon a very timely application of analytics, to determine when somebody is fabricating data or content vs. when they are not. For textual content, it may be possible, for survey questions, possibly more difficult indeed.

I want to add to what have said above that the big data analysis could be actually seen as a solution to this problem. One aspect of the big data analysis is relaxation of the requirement to have structured data, and this implicitly means that the data could be sourced from various, even-incompatible sources into the data warehouse and then into the models. This means that the maintainer of those websites could look for other sources of information about their clients outside their questionnaires.

Also, I would like to mention that even if a client honestly answers to all the questions, there is no guarantee that their behavior does not change after the engagement. An example could be the gym workout culture prior to a successful matching.

Gerard Lynch: Yes, I do remember that paper on stylometry at the EACL 2012. I was actually a Reviewer for that Workshop on Deception Detection. Our lab contributed a paper then as well. This area seems to be under-represented in text analytics thus far, but it has a great potential, to my mind.

Reza Farrahi Moghaddam: So, are you suggesting cross-verification of some kind, then? Wouldn't it be costly (computationally)? Feasible, at all, on a large scale? Especially, if preferences and opinions, as you say, are likely to change. Thanks for the thought, anyway!

And thank you TO ALL who participated in this discussion today! (Any after-thoughts are certainly still welcome!)

Victoria: I was looking at it from the perspective of a company proposing matches (as suggested by 'big data'). They have the broad view of the data and propose matches as their models dictate. Even if an attribute is binary, increasing amounts of data will tell the the likelihood of it being useful. "Do you like horror movies?" is binary, but if that attribute came out on top it's because the majority of people answer it in a way that leads to a match. It doesn't matter whether a small number lie, or whether they all lie, just that it is a consistent indicator of success. I would expect that if half of all people lied and half didn't, then the attribute would not come out on top. (Hmm ... maybe those attributes at the bottom of the list bear a closer look; are those the ones people lie about the most? Maybe those will guide you to the real discriminant factors ie. what people lie about is more revealing of character than what they share.)

From the POV of a user: They are working with small data eg. a couple hundred profile views, tens of dates? Their experience will cause them to adjust their expectations of what they see on other profiles ie. "Yeah, he's probably shorter." As to their own answers, they might try to 'game' the system based on their experience as well.

With emerging several competitive Compute as a Service providers, it would be feasible for almost everybody to perform computer intensive analyses with 'zero CAPEX.'

Also, subsidizes could be requested. Failed matches or delay in matching would have considerable direct and indirect social costs. Therefore, any actor who works toward reducing these social burdens should be eligible to be supported by the public bodies.

In addition, the matching companies could consider to integrate with each other in the form of a consortium or a mega company in order to aggregate and stabilize their revenue while keeping the diversity. This would also enable them to exchange the information toward 'cross verification' and beyond it, i.e., 'behavior modeling.'

In terms of slowing down changes in behavior, one approach could be organized, crowd-based engagement toward preventing behavior degradation:

http://arxiv.org/abs/1110.0436

Of course, there is that age-old saying -- big data or small -- that "the quality of your results cannot exceed the quality of your data." That said, Big Data tends to be not only large but also high quality (messy, high variety, but often very good). For example, if lying is an issue, then it would need to be for a large number of observations, but if it is an issue for a large number of observations, then it can be quantified as a feature of the data (i.e., size of Big Data is an advantage as much as it is a challenge)

The key is that working with Big Data -- irrespective of its quality -- means using algorithms, often several in sequence or ensemble. Cross-validation and metrics (ROC/AUC) then provide feedback about the results of the algorithms individually and in combination. In this way, a "Big Data" set can actually "play against itself" and reveal information such as "a significant subset of users who are lying."

Hi Victoria:Sorry for the belated feedback to your response of a week ago. My use of "lying" in my response to your original post, was based on what the post said -- i.e., it appeared that people were "lying" and that this was negatively impacting match making. I have no idea whether the lying (maybe, "deception" would be a better choice) is benign or not and what the impact if any on the analyses is. Your post suggested that the deception was negatively impacting match making. I agree that there are different types and extents of deception and nuanced interpretations would be necessary, based on context and instance. I was just trying to point out the dissimilarity of the scenario you brought up to the scenario that is commonly addressed in data mining -- detecting fraud in transactions that Ted alluded to. The latter is usually a case of supervised learning with transaction records being marked as fraudulent whereas in your scenario, this is clearly not the case. In that sense, the two scenarios are different and, I think, yours is messier to deal with, if indeed the deception must be dealt with, that is. I say this as others have opined that such deception may be a non-issue, under certain conditions (although which condition prevails (e.g., did a large number lie or only a small number), may also not be easily determined!).

Remarkable terms of big data are volume, velocity and variety

Bid data Analytic based on Latency, Concurrency, access pattern, queries, data scope, end user and technology.

Instance of big data in Retail, marketing and enterprise etc

Manoj Karathiya, thanks for your contribution, but I'm not sure I'm following. I get the 3'V's, but I'm rather talking about the so-called 4th 'V' of the big data - veracity. The question was about potential influence of lying on big data analytics (by the virtue of infidelity of the data). Perhaps you could expand on your thoughts to make it clearer for me what you meant to say.

If everybody is lying, then the regular methods for spotting irregularities wouldn't help. Also, with big data, we don't have a choice but to rely on the data in hand, because there is simply no time to chase for cross checking information.

In my solution (GT data mining) the cheating is treated through identifying the patterns and their characteristics.

In order to mark the lyiers groups I'd be watching for absurd proportions of characteristics which are considered attractive together with typical concealment methods and the ratio of disappointment messages. There could be more or other indicators, depends on what are the input fields in this job. Among the finding there should be lying factors as well. Big data analytics always brings up surprising evidences.

My experience in Big Data research and analysis tells me that every single data analysis scenario in Big Data has its owns pros and cons. The reason is very simple. the people in charge of the technology side, i.e. programmers, data miners should just take care of the technological side and provide the widest and richest data analysis to people, specialist in every single data analysis area. In other words in order to avoid data analysis bias every single data analysis must be done bu people specialized in that area. For instance, a social study can be done easily and with more effective results if a Sociologist specialized in that single area is called to the team to perform the data analysis.

Let's leave technology related people deal with the technical area and experts in specific areas deal with the data analysis.

Dear ,lying even its few ,may affect the data analysis ,and lead to false results .