I'm divorced with (two superrrrrrr awesome) children. My ex-wife and I get along great. I have a superb set of close friends. Life's good. But I would love to find an amazing partner and remarry, and possibly even have more children together. I understand that every person has different ideas of what is important to their own sense of happiness and fulfillment. For me, being in a great marriage is pretty high on the list of things that will have a deep and prolonged impact on my personal growth and satisfaction in life. So, the best way to meet people is to ...meet people.

With the data available, I had questions about my own dating life, such as the importance of factors in my own personal affinity for dating certain people versus others, and which aspects of my connection with someone seemed to have the strongest impact for determining a possible successful pairing. What insights could I possibly gain that might help me identify better matches? Is there something I could do better in general?

What follows below is my own personal life laid bare. Some results are a bit surprising. Others are probably expected. It's holding my search for a partner up to the data science mirror and seeing what is staring back at me in the reflection. Along the way I show how basic data science principles are used and how the results they provide may or may not be directly actionable. And last, there's a fair bit of poking fun at myself and the overall silliness that is trying to use data science to find a perfect lifelong match. Love and life is much messier than a couple variables' relative percent importance, of course. But it's fun and interesting nonetheless.

1. The Data-base Breakdown
   • Feature Definitions
2. Super Model
   • Data Prep
3. The First Time
   • Feature Importance
   • Relative Correlations for Age vs Attraction
   • Logit Coefficients
4. Age is Just a Number
   • Binning
5. Second Run
   • Feature Importance
   • Dare to Nair
   • Feature Importance Without Hair
   • Height
   • Intellect and Humor
   • Bad Dates
6. Politics
   • Voters by County
7. Chemistry
   • Feature Importance
   • You Really Really Like Me
8. The Friend Zone
   • Likes and Friends
9. Attitude is Everything
   • Attitude Counts
   • Attitude and Income
10. Check Please
11. References

My dating database had the following variables:

_{Table 1: The variables I assigned to each person in the database.}

Let me quickly clarify how certain variables were assigned.

• Attraction: Physical attraction only.
• Intellectual Connection: Basically the mental equivalent of attraction.
• Humor: How well our senses of humor aligned; did I laugh a lot with this person?; did they get my amazingly bad dad jokes?
• Chemistry: The level of physical chemistry.
• Life Attitude: This is an appraisal of the person's general attitude in life. This applies for men and women equally, of course. I just happen to be interested in women, though I'm sure every woman I've met could slot the men they've dated in these exact same categories!
  • Positive: Generally optimistic, able to handle changes (even minor things like change of plans on a date) well. If the waiter makes a small mistake on their order, they just take off the unwanted food or, if necessary to address it, do so politely and don't mention it again. Has a good view of the "big picture."
  • Neutral: Depends on the mood or situation. More of a 'nondescript' attitude.
  • Negative: Generally pessimistic. Sees obstacles as larger than they are, lets little things bother them. Would say things like "I don't want to go to [Place X] because they messed up my order one time."
  • Complainer: A specific sub-set of Negative, in that I found the person to be routinely not just negative in a "pessimistic about life" sense, but to explicitly complain about many things and people. If the waiter made a mistake with their order, they would let it impact their ability to enjoy the meal, be fairly snarky to the waiter while demanding a correction, and finally -- and this is the coup de grâce -- they would mention it as the first thing when someone asked about the date. Like, who cares? Your order wasn't perfect. Wow. Big deal. Get a life. How on earth these people can find any happiness in life, I will never know.
• Income: Since I don't have the tax returns of everyone in this database, I don't know exactly where they fall in income. Things such as vocation, if they owned a home, neighborhood they lived in, type of car, and cost of hobbies are the types of things I used to judge a person's income. It is imperfect but is likely representative to a fair degree. So I made the following rough annual income approximations.
  • Low: less than $40,000
  • Medium: between $40,000 and $80,000
  • High: above $80,000
• Second Date: Whether I saw this person a second time or not. On the surface this seems like it would be an easy winner for most predictive variable -- if I "liked" someone then surely I'd see them again, right? Well, life isn't always so cut-and-dried. Sometimes circumstances change. Sometimes you like someone that doesn't like you. Sometimes you met only by chance to begin with. Life can be fickle. Hence the inclusion of this feature.
• Did I Like This Person?: At the end of the day, whether I would consider full-on dating this person or not. Ultimately, this is the target of the whole endeavor -- to find someone you'd like to date and, in my case, marry.
• Could We Be Friends?: If I didn't feel like the connection with this person was right for dating, could we still be moderately good friends, including socializing platonically?

My main interest in this mini-project was simple: to see which variables had the strongest relationship to predicting whether or not I liked someone enough to consider dating them. For this I used a Gradient Boosted Classifier from XG Boost. I intentionally overfit the model (500 trees, tree depth of 4, and learning rate of 50%) because I am not interested in building a predictive algorithm using the limited variables I have. The primary issue in doing so is that some of the variables such as attraction, humor, intellectual connection, chemistry, politics, and second date can only be answered fairly after I've met someone. Thus this is a descriptive project and not an inferential or predictive one.

The final rankings of which features are most important in deciding whether or not I "liked" someone are given by the .feature_importances_ attribute of the fitted model. These are determined by the amount of information gained if using entropy as the metric to evaluate successful node splits in a tree, or from the magnitude of increase in node purity if using the Gini coefficient instead.

There were three stages to data prep in this project.

1. Dummy the categorical and binary variables.
2. Address the Age over-importance issue (see below).
3. Add Chemistry feature.

Fo the first run of the model I had not yet created the Chemistry variable, or as they say in Russia, there was no "chim-ih-stree." It turns out this mattered. We'll come back to this in our second run below. Here are the top 10 features for the (overfit) GBC model:

_{Table 2: The top 10 features from my initial run, without Chemistry, and without any data prep apart from dummied variables.}

My initial reaction to this result was one of surprise. Age? Really? That's the overwhelmingly most predictive feature for whether or not I liked someone? That sure didn't feel right, so I did a quick visual inspection to see if this made sense by looking at the basic Pearson R² correlation between Age and a binary version of the Like_This_Person? target, as well as the same for Attraction.

_{Figure 1: Comparison of Age and Attraction correlation to 'Liking' someone.}

Visually the result doesn't seem to support the feature importance results above. The slope for Age is much flatter than that for Attraction. This is not a definitive result (for one, because we shouldn't use a linear regression in this situation); it's just a basic litmus test to see if things are in the right ballpark. This plot suggests that we are not.

Curious, I wanted to find the basic correlation between Age, Attraction, and Like_This_Person? to further support that the initial feature importances seemed askew. Since our dependent variable of Like_This_Person? is binary, a standard OLS linear regression is not advised. So, using Sci-Kit Learn's LogisticRegression module (C=1.0, penalty='l1'), the following coefficients were calculated for the two variables in question.

_{Table 3: Logistic Regression coefficients for Age and Attraction, showing a greatly diminished importance for Age.}

This isn't meant to be a comprehensive logistic regression report using an optimally-tuned Logit model, but rather a guideline as to the relative importance of each feature. So both graphically (Figure 1) and mathematically we are seeing that Age does not appear to be particularly important. But our (more advanced) gradient boosted model said that Age was not just important, but the most important feature by a mile. What gives?

Well, this is where understanding what's going on "under the hood" for the models we are using shows its importance. When using a decision tree-based model, of which gradient boosting is an ensemble version, there is a specific susceptibility for inflated feature importance reporting if a feature can take on a large range of unique values relative to other features in the dataset. ¹ As the decision tree makes a split, many of the features in this dataset are binary (dummied) and provide little difference for the model to measure.

Looking back at Table 1, we see the range of values that Age can take is double any other numeric feature! The model is getting a unique value on a split of Age at around twice the rate it is for the other numeric columns and an order of magnitude more than the binary ones. (Credit to my friend Kayla for collaborating on this point). Thus, it seems safe to say that the reported importance of Age is a relic of the database's construction and is not indicative of its true value in determining if I will "like" someone.

Clearly we need to fix this error. A first thought might be to simply dummy the Age variable as we did with the categorical ones at the beginning of the project. However, this would result in two problems. First, this would create N = df['Age'].unique().shape[0] = 20 dummy columns. Second, we risk completely losing the essence of the Age variable. With 20 columns each only representing a single age, it is incredibly unlikely that any single age will register as "important." This is due to there being relatively few entries of any given age, so the each split on a given column of age is unlikely to include a "yes" for that age -- there just aren't going to be a majority of people in this database that share the exact same age.

Further, what if the model said "Age 31" is the most important age? What would that really even mean? Are we saying that I am only going to "like" someone if they're 31? If I met that same person a week before or after they were 31 years old, would I not like them equally as much? It appears that a basic dummying of the Age column is not going to resolve our issue. Fortunately, there is an easy and interpretable solution.

Instead we will bin our Age data. As anyone who's made a basic histogram in MS Excel knows, you sort data into groups based on different cutoffs. For our Age data, we'll bin into quintiles (fifths). This way we can have a group structure of "very low, low, average, high, very high" for the ages in the database. There is one catch, though, and that is the count of people in each group. Are we going to make the group cutoffs based on percentile ranking or percentile membership?

For a quintile binning, the ranking approach simply takes all the data points in the bottom 20th-percentile (the first quintile) and puts them in the "very low" bin, data points in the 20th- to 40th-percentile (the second quintile) and puts them in the "low" bin, all the way up to data points above 80th-percentile (the fifth quintile) being shoved into the "very high" bin. If your data is distributed roughly evenly, this is fine, else it is vulnerable to outliers causing greatly imbalanced bins.

Let's pretend there was a person of age 55 in this database. Using a percentile ranking, this person would rank in the 100th-percentile and the nearest person below them would be in, oh, the ~70th-percentile. As a result, the top bin of "very high" would have only one person. What statistical conclusions can be drawn from a lone data point? (Hint: none). In Pandas, this method is pd.cut().

The other option is to bin the data based on count. We simply count the number of data points and put the bottom scoring fifth into the "very low" bucket, regardless of if they're all actually below the 20th-percentile in their value. We repeat this for each bracket, up to taking the highest fifth of the dataset and putting it into the "very high" bucket. This approach ensures (as close to possible) balanced bins and is impervious to extreme outliers. This is the route we will follow in this model. In Pandas, this method is pd.qcut() (which stands for "quantile cut" and can divide your data into any N number of divisions).

With our binned Age variable, the feature importance results are notably different and, anecdotally speaking as the person who has met the people in this database, much more accurate.

_{Table 4: Feature importance results after binning the Age variable.}

...and what a difference proper data preparation has made! By binning the wide-ranging Age column we have let the other features in the model "breathe," giving them a chance to show their relative importance much more accurately. The results themselves are somewhat interesting.

First, note that the two types of hair are basically equal, showing that I don't seem to have a preference based on hair. This is a fun finding for me, because I would have absolutely said that I prefer "dark" hair over blonde hair, but all things equal it seems that has very little to do with whether or not I actually consider someone as a good candidate to date long-term. In a sense, this makes both features redundant -- every woman I have dated has had hair, so we aren't learning anything by keeping it in the model at this juncture.

Before we proceed without Hair, it's worth noting the breakdown of the category. In the dataset there are 27 Blondes, 24 Brunettes, and 1 Red. Further backing the notion that I don't seem to have a real preference is that the mean Attraction for both brunette and blonde is nearly identical at 5.88 and 5.90, respectively. Again, this is news to me! It's definitely an experience to see your own life be churned out in data and tell you something you might not have known about yourself. This is a positive result, because it means there's one less category to care about when searching for a lifelong partner.

Someone inexperienced with data science might also conclude that the fact there is only one redheaded person in the database means I am not attracted to women with red hair. But, as always, we have to dig one level deeper to see if we can properly contextualize the results we get. In this case, there are 52 total people in this database. One has red hair. That's just a, uh, hair beneath 2% of the total database. Well, a quick Google search gives us the context we are looking for.

Care to guess what percentage of the world population has red hair? ...if you said 2%, you win the Daily Double! So, the occurrence of a lone red headed person in the database is in line with what we'd expect given the size of the database (though, it would vary by country. The USA seems to be a fairly close representative of world red haired ratios). Thus, we can't say anything conclusively whatsoever about my preference for or against red hair!

Now, let's run the model with Hair removed altogether.

_{Table 5: Feature rankings with the Hair column removed.}

It looks like the biggest beneficiary of removing Hair was Height, as its importance increased by around 14% relative to the rest of the database, keeping it in the pole position. Having Height as the most important feature presents us with an excellent opportunity to point out that tree-based feature importances don't tell us whether a higher or lower value of a given variable helps predict the desired outcome, only that one of the two does. In the case of Attraction, the obvious conclusion is that people I was more attracted to are going to be more "liked" by me and considered a possible fit for long-term dating.

But we have no such obvious conclusion with Height. It could mean that I routinely "liked" taller women, or that I routinely "liked" shorter women. We don't know from this table alone. One way to find out would be to do some exploratory data analysis (EDA) on the dataset and see if we could note the trend ourselves. Yet another option would be to ask me, since I probably have an inkling of which type of person, taller or shorter, I seem to respond favorably to. I would think this hunch should be especially evident to the subject (I) since Height appears to be such a salient factor. Given this, let's see if we can construe a viable answer from a quick and dirty plot of the data.

_{Figure 2: Heights of everyone in the database. 5'7" is the midpoint of the range.}

Judging from this plot of heights, it would appear that shorter heights correlate to my liking someone. As the person whose dating history this database is referencing, I can confirm that indeed I seem to find myself naturally more drawn to shorter rather than taller partners.

Following Height and Attraction are Intellectual Connection and Humor. It makes sense that these features round out the top four. When we think of choosing a partner, obviously we would like to be attracted to them (studies have shown that men tend to value attraction slightly more than women, to no one's surprise). But right along with attraction we want someone we can "connect" with (see: hold multiple conversations that are actually stimulating, teach me something, or open my eyes to something I had never realized before) and someone who can make us laugh or that we can laugh with. I surely would have said as much before making this project, so it's pretty neat to see the data bear it out!

We should quickly touch on Second Date (No). Despite the possibilities for undesired absences of second dates mentioned in the intro, it appears that most of the time a second date doesn't occur has probably been either my choice or a mutually agreed upon lack of interest. I say this because of the feature's appearance here. Again, these feature importances don't tell us directionality, but with the category being split into a (No) sub-set, we can conclude that the absence of this feature (meaning that there actually was a second date, since the "No" column would be empty) correlates to my "liking" someone. This makes sense.

Politics Left is a bit of a surprise, personally. Again, just from its reported importance we are not told whether have more or less of this variable predicts if I will "like" someone. I'm a registered independent and find radical political views on either side to be a moderately off-putting trait in a potential partner. I find strict ideology in general is a suffocator of innovation and critical thinking. So my personal political affiliations don't offer much of a roadmap, here. (If I belonged to one party we could deduce whether this variable, Politics Left, would be considered a good or bad thing. But I don't, so we can't). I can't help but wonder if there's not something more to this, else I might just chalk it up to being fake news.

_{Table 6: Mean summary values for each political affiliation in the database.}

Interesting. There isn't anything that jumps off the page and screams "I clearly favor women who have a certain political view." Women leaning left politically edge out those leaning right in only one of the four variables, Intellectual Connection. Conversely, conservatives come out on top in Attraction, Humor, and Chemistry. But none of these "leads" are particularly resounding. Humor is almost deadlocked, and the other three categories are roughly mirror images of each other.

There are, however, two pronounced observations to be made. First, the number of democrat women more than doubles that of republicans. Again, someone eager for a quick conclusion might read that as suggesting I like to date liberal women twice as much as conservative women. But, we can do better than such surface level takes. The key element to any demographic sample like this is: where was it taken? The entirety of this dataset was taken during my time living in or near Boulder, CO and the surrounding Denver metro area. It is not surprising then that we'd expect to see a strong prevalence of democrats in this sample.

The second point to note is that I apparently mesh very well intellectually with independents. The sample size is too small to draw any sound result, but averaging over nine-out-of-ten in both intellectual and humor connection for four people is compelling. This makes perfect sense -- I, myself, am an independent, so I would expect that I'd get along best with those who are also independent politically. It's just so starkly portrayed here that even I chuckled. Data for the win (FTW)!

Out of curiosity, I looked up voter registration statistics for the counties I've lived in or that neighbored me.

_{Figure 3: Voters by neighboring counties in CO, May 2017.}

Well, the interesting point to me as an independent voter myself is how numerous independent voters are! In fact, it looked like this might be the result of an error, but a hasty Google search showed that 37% of all Colorado voters are independent/unaffiliated! ² This stunned me. I assumed it was around 15%. Regarding the ratio of left-to-right political representation in the database, I've lived mostly in either Boulder or Broomfield counties. It is not surprising that a higher percentage of my dates have been with liberal women, then. It would be foolish to conclude that I have a predilection for dating women with liberal views just based on the number of such women I have dated when statistically it is right in the range we'd expect based upon geography.

Only two counties, Weld and Larimer, have more registered republican than democrat voters. Both counties are at the northern border of the state and are home to much of the state's oil and gas industry. In the other six counties republican voters are in last place (Adams, Arapahoe, Boulder, Denver) or tied for last (Broomfield, Jefferson). The selection pool is just smaller for conservative-minded partners in this area of the state. It's that simple.

However, Figure 3 also suggests that I have had a dearth of independent voters in my dating experiences, which when we see how highly they rate in Intellectual Connection and Humor -- the third and fourth most important features in the "hairless" run of the model! -- is something of a real shame. In fact, independent voters are the leading registration alignment in five of the eight counties! Absolutely shocking results to me. If I was to glean any advice going forward from this mini-study, it would perhaps be to make an effort to try to meet more independent voters.

In a moment of inspiration after fixing the Age binning issue and investigating political and hair color trends, I realized I had not included a potentially important variable: Chemistry. Anyone who has been in a serious relationship knows that a large measure of appeal and happiness stems from chemistry. A general quality of "do this person and I 'click' with each other physically?" is a core component of the glue that holds a relationship together. In some ways, this is a more implicit aspect than is Humor, for example. As such, I added it to the database and ran the model again.

_{Table 7: Feature importances with Chemistry included. Note that Hair is still excluded.}

There's just no overriding biology, is there? Chemistry absolutely obliterates every other factor. When I initially set out on this fun little endeavor, I was expecting (hoping?) to find a result that had Intellectual Connection at the top. It is, after all, a trait I believe contributes very strongly to a long and happy relationship. And while that exact statement may be true -- that an intellectual rapport between you and your partner is necessary for sustained fulfillment -- it seems clear that it isn't the factor that's going to get you into a long relationship in itself. The numbers don't lie, and they speak a resounding oration in this discussion.

If I have good chemistry and feel comfortable with someone, then I will probably "like" them enough to want to continue seeing them. If we don't have a comfortable connection in that aspect, then it seems clear that the concept of being with that person for the long haul is a no-go.

In real chemistry we say "like dissolves like," referring to solvents. And I suppose that's largely what's at work here. If there is a productive reaction together, then you both feel alike and this feeling (a potentially admittedly rare one) can dissolve away other impediments, such as political differences or slightly diverging senses of humor, allowing a successful relationship to precipitate out. The differences in other areas likely act as a product loss -- you aren't going to achieve maximum possible solvency if you intermittently come to loggerheads over politics and humor -- but you are at least miscible enough to overcome the activation energy required to start the relationship in the first place.

Okay, that's the last of the chemistry-laden metaphors. I promise. All things equal, I think this makes sense, somewhat disappointing though it may initially be. No matter how attracted you are to someone, no matter how much you laugh, or how lockstep you are with each other on politics or world affairs, if you are romantically involved with each other and it doesn't feel right, there's just no realistic chance you would want to continue seeing that person in a romantic role. Basic biology wins again.

I guess I wouldn't be here to write this and you wouldn't be here to read it if Chemistry wasn't a highly powerful driver. So, in a way seeing Chemistry at the top of the list is an affirmation that I'm wired like all my ancestors before me, and this is a vestige of epochs of life that saw our species' survival play out strictly as a game of numbers.

A fairly common result when dating is to meet someone who you don't have much romantic chemistry with but who you get on with well enough to be platonic friends. In this recent dating history of mine there were 52 people in the database, 21 of which I "liked" and 25 I wanted to remain "friends" with. That's a reasonable trend -- it's probably easier to find someone to be friends with than it is to date -- but it also includes a subtle difference that needs to be highlighted: one is not necessarily a subset of the other!

We would think that liking someone enough to date means that you could also be friends, and this is often true, but it isn't always so. Due to myriad factors it is possible to think of someone as a possible candidate for dating while not thinking of them as a friend necessarily. I observed this very phenomenon in my personal dating history. Of the 21 people I "liked" enough to date, only 16 were labeled as someone I would want to be "friends" with. I made a Venn diagram to help illustrate this result.

_{Figure 4: Venn diagram of the people I "liked" and the people I felt would make a good friend.}

Sixteen of 21 people represents 76.2% -- almost exactly three out of four. That means that one out of four people I felt I could date were people I didn't really want to be "just friends" with. I find this result somewhat surprising! I wonder if the ~25% of people I would date but wouldn't care to be friends with would have made horrible relationships? How could you reconcile that determination with your other friends or family?

[To friends]
ME: Oh, yeah, Rapunzel -- I'm just absolutely crazy about her.
FRIEND 1: Really? That's great to hear. Always nice to find a real friend that you also click with romantically.
ME: Whoa, hold on now. I didn't say anything about being friends.
FRIEND 2: Uh, whaddaya mean? You just said you were crazy about this girl? You guys have been together for a while now. That's kind of the whole point of a relationship, getting to be intimate with someone that is also your best friend.
ME: Yeah, but I wouldn't actually consider her a "friend." I would date her, but I don't think we are a good match as friends. We aren't at that level, if you know what I mean?
FRIENDS: [in unison] No, we have absolutely no idea what you mean. In fact, just stop talking. Thanks.

_{Table 8: Averages for the numerical features for people I liked romantically versus the ones I didn't as well as the ones I felt could be good friends versus the ones I didn't.}

The grouped averages for the four results of "liking" and "friending" (work with me) show the inflated importance of Attraction and Chemistry in determining whether or not I felt someone would be a potential dating match compared to just making a good friend. Conversely, we also see an increase in the role of Humor in deciding if someone would be a viable friend over being someone I'd like to date. Intellectual Connection remains about the same for both.

Interestingly, however, we see that while Attraction and Chemistry show no real difference between not viewing someone as a possible friend, Intellectual Connection and Humor both are markedly lower for people I judged to not be a prospective friend compared to those I simply felt wouldn't be right to date. In other words, if I didn't feel good chemistry or attraction to someone then I wouldn't be keen on dating them, but I could very well consider them still a friend. However if we didn't have a fair mental or comedic connection, then I felt there was little to tie us together as even just friends.

The Logit coefficients for which variables most correlates to labeling someone a valid "friend" bear this out, too, with Humor and Intellectual Connection leading the way at .692 and .560, respectively. Attraction brings up the rear at .261, showing friendship in my view has little to do with physical appearance. (Yay me?)

We don't need an NIH-funded Harvard study to tell us this, of course. But it is neat to see it borne out in my own dating history. By and large, I have to be on the same page intellectually or comedically if I want to be reliable friends with someone. If I am attracted to them and have a romantic chemistry, then I appear to be willing to "trade in" a bit of mental rapport and mirth for the even more sought-after and more elusive true intimate connection. I suspect my experience with this echoes that of most people my age and younger, for better or worse. My father is 65 and I am certain he wouldn't value, intentional or subconsciously, a great romantic connection with someone at the same level I do. And when I am 65, I almost certainly won't either. All a part of life!

On a lark, I had a question come up that I wanted to poke around on. "Does personal attitude correlate with personal income," I wondered? While my little dating dataset would by no means provide any sort of conclusive answer to this, I thought it might be fun to see how the numbers played out. Now, recall that I had no clue what the answer would be, nor did I even consider this as I was making the database. For clarity, let's take a look at the breakdown of the Attitude category.

_{Table 9: The number of people in the database that have a given income at a given attitude. The total categorical counts for Attitude are Positive 17, Neutral 13, Negative 13, Complainer 9.}

We see a fairly general trend -- more people in low than high income brackets -- with a few more Positive than the rest, and Complainer has the fewest (which is a good thing). So, if there's no real trend then we should have a roughly equal even breakdown of incomes. If, however, there is some underlying component of attitude that might correlate with personal income, it would likely show up to some degree in a plot.

_{Figure 4: People's income as grouped by what I found their personal attitudes to be. Look at the blue bars.}

This is my favorite result from this entire project! Look at the Positive (blue) bars! I've always believed that a positive attitude in life is going to net better outcomes. You handle stress better. You handle short-term obstacles better. You handle your failures better. You handle others' failures better. You see challenges as ways to learn and grow instead of likely ways to falter. This chart tickles me pink.

Do recall again that I made this database without any notion of these results. This question wasn't even one I had conceived until I began writing my lecture on using this database as a means to demonstrate basic concepts in Pandas. And we should also note that something like "personal attitude" is inherently a malleable concept. What I consider "neutral" someone else may peg as "positive." By nature I am inherently strongly positive and I naturally seek this quality in others, especially a partner. So, to see this breakdown implies a few things, none of which are conclusive of course, but all of which are interesting.

First, if there is a link between personal attitude and financial success, these results suggest my labels of people's attitudes are reasonably accurate.

Second, these groupings may imply that either I am slightly mislabeling people's incomes or that I have had a larger than average share of low income people in my dating life. I feel very strongly that overall the income labels I assigned are representative. That in itself might be more a reflection of the partners I was encountering than anything else, I suppose. Still, if I got the income category 90% right, that would mean that up to five people were misclassified, which could change the layout of Figure 4.

Third, these results would seem to suggest (only suggest) that there could indeed be a relationship between personal attitude and financial success. The most negative group, Complainer, has eight of its nine entries (88.9%) in the low bracket. The lone exception is in the medium group, meaning the Complainer group is the only one that doesn't have anyone in the high income group. See for yourself:

_{Table 10: Percent of each Attitude group in a given income bracket.}

This is really compelling! Look at any income bracket and go down the column. The Positive attitude group has the best score in all the brackets, with the lowest percentage of low incomes and the highest in medium and high. Neutral and Negative attitudes are identical, perhaps hinting that there isn't a ton of difference between the two, or that I could have labeled some members of these two groups incorrectly in either Attitude or Income. Or it could just a random chance, as the sample size for each is only n=14.

Nothing conclusive can be drawn from these results, of course, but they didn't do anything to dissuade me from my belief that a positive approach to life leads to the best outcomes in life.

Thanks to this project I learned that biology is driving so much more of my partnering choices than I would have known, or maybe even cared to know, otherwise. I also learned that Colorado has seen a swell in the number of independent voters in the last decade, now totaling a whopping 37% of all registered voters (this despite the inability to vote in any caucus!), and Denver as a whole has become more blue than purple in that time as well (it was significantly more split when I first moved here). So as long as I am in the Denver metro area, my choices in possible partners are probably going to stay in the "more left and equally red headed" as other parts of the country. And using just people from my own personal dating history, I was surprised to elucidate a possible trend in the power of a positive attitude in affecting one's own income.

In the end, I don't know that there is a silver bullet to finding the right match for any one person -- love and connection are in some ways a chance result from the confluence of many confounding factors. Were you in a stable place in life when you met? Were you able to emotionally open up to someone? Did you have the time or the income to devote to a real relationship? Did you even know what kind of relationship you wanted? ..and a million other things. This project helped me realize I'm a pretty standard person, man or woman, in that I'd like to have a good mental and physical connection with someone before I'd consider them a possible long-term match. Who wouldn't say that about themselves?

Here's hoping the next dating database I make will be one of the dates my wife and I traveled the world and did awesome stuff together (get it? Ha? Ha? Ha?). Until then...

1. https://en.wikipedia.org/wiki/Information_gain_in_decision_trees#Drawbacks
2. http://coloradoindependentvoters.strikingly.com/