Exercise: �2 Mutual Information
In this exercise �you'll identify an initial set of features �in the Ames dataset �to develop using mutual information scores �and interaction plots.��Run this cell to set everything up!
注目する特徴量の選び方
本文で使用した２つの関数
家の価格を予想したい
The Ames Housing dataset was compiled by Dean De Cock for use in data science education. It's an incredible alternative for data scientists looking for a modernized and expanded version of the often cited Boston Housing dataset.
MoSold 売れた月
YearBuild 建築年
Screen Porch 網戸付きベランダ
Based on the plots, which feature do you think would
have the highest mutual information with SalePrice?
The Ames dataset has seventy-eight features –
a lot to work with all at once!
Fortunately, you can identify
the features with the most potential.
Use the make_mi_scores function
(introduced in the tutorial)
to compute mutual information scores for the Ames features:
78個の特徴量から
重要な特徴量を選びたい
2) Examine MI Scores
Do the scores seem reasonable?
Do the high scoring features represent things you'd think most people would value in a home?
Do you notice any themes in what they describe?
Some common themes among most of these features are:
These are all the kinds of features you'll commonly see in real-estate listings
(like on Zillow), It's good then that our mutual information metric scored them highly.
On the other hand, the lowest ranked features seem to mostly represent things that are rare
or exceptional in some way, and so wouldn't be relevant to the average home buye
and counts like FullBath and GarageCars
Foundation（ファウンデーション：基礎土台）
These are all the kinds of features you'll commonly see in real-estate listings
(like on Zillow), アメリカの大手不動産情報サイト
It's good then that our mutual information metric scored them highly.
On the other hand,
the lowest ranked features seem to mostly represent things
that are rare or exceptional in some way,
and so wouldn't be relevant to the average home buye
不動産の「SF features」は、「Special Features」の略で、物件の付加価値を高めるような特徴や設備のことです。具体的には、キッチンやバスルームのグレードアップ、広々としたバルコニー、最新のセキュリティシステムなど、他の物件とは異なる魅力的な要素を指します。
In this step you'll investigate possible interaction effects for the `BldgType` feature. This feature describes the broad structure of the dwelling in five categories:
Bldg Type (Nominal): Type of dwelling>
1Fam Single-family Detached
2FmCon Two-family Conversion; originally built as one-family dwelling
Duplx Duplex
TwnhsE Townhouse End Unit
TwnhsI Townhouse Inside Unit
The `BldgType` feature didn't get a very high MI score.
A plot confirms that the categories in `BldgType` don't do a good job of distinguishing values in `SalePrice`
(the distributions look fairly similar, in other words):
住宅
Still, the type of a dwelling seems like
it should be important information.
Investigate whether BldgType produces a significant interaction
with either of the following:
GrLivArea # Above ground living area
MoSold # Month sold
地上面積から住宅価格を
予測するのには
使えそう
住宅タイプに分ければ
販売月から住宅価格を
予測するのには
使えなそう
3) Discover Interactions
From the plots,
does BldgType seem to exhibit an interaction effect
with either GrLivArea or MoSold?
交互作用 2つの因子が組み合わさることで
初めて現れる相乗効果のこと
相乗効果 複数の因子が作用するとそれぞれの独自の効果の総和よりも大きい効果を表すこと
The trends lines within each category of BldgType are clearly very different, indicating an interaction between these features.
Since knowing BldgType tells us more about how GrLivArea relates to SalePrice, we should consider including BldgType in our feature set.
The trend lines for MoSold, however, are almost all the same. This feature hasn't become more informative for knowing BldgType
3) Discover Interactions
From the plots,
does BldgType seem to exhibit an interaction effect
with either GrLivArea or MoSold?
交互作用 2つの因子が組み合わさることで
初めて現れる相乗効果のこと
相乗効果 複数の因子が作用するとそれぞれの独自の効果の総和よりも大きい効果を表すこと
The trend lines for MoSold, however, are
almost all the same. This feature hasn't become more informative for knowing BldgType
Let's take a moment to make a list of features we might focus on. In the exercise in Lesson 3, you'll start to build up a more informative feature set through combinations of the original features you identified as having high potential.
You found that the ten features with the highest MI scores were:
OverallQual 0.581262
Neighborhood 0.569813
GrLivArea 0.496909
YearBuilt 0.437939
GarageArea 0.415014
TotalBsmtSF 0.390280
GarageCars 0.381467
FirstFlrSF 0.368825
BsmtQual 0.364779
KitchenQual 0.326194
対象不動産の周辺にある、生活や商業活動など、人間活動に関連してまとまりを示しているエリアを指します。
Do you recognize the themes here?
Location, size, and quality.
You needn't restrict development to only these top features, but you do now have a good place to start.
Combining these top features with other related features, especially those you've identified as creating interactions,
is a good strategy for coming up with a highly informative set of features to train your model on
Keep Going 次の節へGO
and learn
what kinds of transformations
different models are most likely to benefit from.