Appendix-Datasets.qmd

---
output: html_document
editor_options:
  chunk_output_type: console
fig-cap-location: margin
---

# Datasets {#sec-datasets}

```{r}
#| include: false
knitr::opts_chunk$set(fig.align = "center")
```

```{r}
#| echo: false
#| include: false
#| message: false
#| warning: false
reticulate::use_virtualenv("r-keras")
library(tensorflow)
library(keras3)
tf
```

You can download the data sets we use in the course <a href="http://132.199.73.15:8500" target="_blank" rel="noopener">here</a> (ignore browser warnings) or by installing the EcoData package:

```{r chunk_chapter8_0, eval=FALSE}
devtools::install_github(repo = "florianhartig/EcoData", subdir = "EcoData",
                         dependencies = TRUE, build_vignettes = FALSE)
```

## Machine learning pipeline / workflow

![Machine Learning pipeline](images/pipeline.png)

## Titanic

The data set is a collection of Titanic passengers with information about their age, class, sex, and their survival status. The competition is simple here: Train a machine learning model and predict the survival probability.

The Titanic data set is very well explored and serves as a stepping stone in many machine learning careers. For inspiration and data exploration notebooks, check out this <a href="https://www.kaggle.com/c/titanic/data" target="_blank" rel="noopener">kaggle competition</a>.

**Response variable:** "survived"

A minimal working example:

1.  Load data set:

```{r chunk_chapter8_1}
library(EcoData)

data(titanic_ml)
titanic = titanic_ml
summary(titanic)
```

2.  Impute missing values (not our response variable!):

```{r chunk_chapter8_2, message=FALSE, warning=FALSE}
library(missRanger)
library(dplyr)
set.seed(123)

titanic_imputed = titanic |> select(!any_of(c("name", "ticket", "cabin", "boat", "home.dest")))
titanic_imputed = missRanger::missRanger(data = titanic_imputed |>
                                           select(-survived), verbose = 0)
titanic_imputed$survived = titanic$survived
```

3.  Split into training and test set:

```{r chunk_chapter8_3}
train = titanic_imputed[!is.na(titanic$survived), ]
test = titanic_imputed[is.na(titanic$survived), ]
```

4.  Train model:

```{r chunk_chapter8_4}
model = glm(survived~., data = train, family = binomial())
```

5.  Predictions:

```{r chunk_chapter8_5}
preds = predict(model, data = test, type = "response")
head(preds)
```

6.  Create submission csv:

```{r chunk_chapter8_6, eval=FALSE}
write.csv(data.frame(y = preds), file = "glm.csv")
```

And submit the csv on <a href="http://132.199.73.15:8500" target="_blank" rel="noopener">http://132.199.73.15:8500</a>.

## Plant-pollinator Database {#sec-plantpoll}

The plant-pollinator database is a collection of plant-pollinator interactions with traits for plants and pollinators. The idea is pollinators interact with plants when their traits fit (e.g. the tongue of a bee needs to match the shape of a flower). We explored the advantage of machine learning algorithms over traditional statistical models in predicting species interactions in our paper. If you are interested you can have a look <a href="https://besjournals.onlinelibrary.wiley.com/doi/full/10.1111/2041-210X.13329" target="_blank" rel="noopener">here</a>.

```{r chunk_chapter8_7, echo=FALSE}
knitr::include_graphics("./images/TM.png")
```

**Response variable:** "interaction"

A minimal working example:

1.  Load data set:

```{r chunk_chapter8_8}
library(EcoData)

data(plantPollinator_df)
plant_poll = plantPollinator_df
summary(plant_poll)
```

2.  Impute missing values (not our response variable!) We will select only a few predictors here (you can work with all predictors of course).

```{r chunk_chapter8_9, message=FALSE, warning=FALSE}
library(missRanger)
library(dplyr)
set.seed(123)

plant_poll_imputed = plant_poll |> select(diameter,
                                           corolla,
                                           tongue,
                                           body,
                                           interaction)
plant_poll_imputed = missRanger::missRanger(data = plant_poll_imputed |>
                                              select(-interaction), verbose = 0)
plant_poll_imputed$interaction = plant_poll$interaction
```

3.  Split into training and test set:

```{r chunk_chapter8_10}
train = plant_poll_imputed[!is.na(plant_poll_imputed$interaction), ]
test = plant_poll_imputed[is.na(plant_poll_imputed$interaction), ]
```

4.  Train model:

```{r chunk_chapter8_11}
model = glm(interaction~., data = train, family = binomial())
```

5.  Predictions:

```{r chunk_chapter8_12}
preds = predict(model, newdata = test, type = "response")
head(preds)
```

6.  Create submission csv:

```{r chunk_chapter8_13, eval=FALSE}
write.csv(data.frame(y = preds), file = "glm.csv")
```

## Species distribution modeling (SDM) {#sec-sdm}

This data set is based on the <a href="https://www.kaggle.com/competitions/geolifeclef-2025" target="_blank" rel="noopener">GeoLifeCLEF 2025</a> challenge and poses a multi-species distribution modeling task: for a set of geographic sites (which we call *communities*), predict which of 448 plant species are present, using environmental predictors that describe each site. Unlike the other data sets, the response is not a single column but a **presence-absence matrix** — one binary label per species for every community.

**Response variable:** presence/absence of each of the 448 species (a community $\times$ species matrix).

**Predictors:** environmental variables characterizing each community — e.g. bioclimatic variables (`Bio1`–`Bio19`, such as mean annual temperature and precipitation), soil properties (e.g. SoilGrids pH), elevation, and similar geographic layers.

### Data structure

For both the training and test communities you get two tables:

- an environmental table (`env_df`) — one row per community, one column per environmental predictor;
- a presence-absence label table (`labels`) — a community $\times$ species matrix of 0/1, with species in the columns (named `sp_9`, `sp_53`, `sp_96`, …).

The test set contains **2000 communities** and **448 species**. The labels of the test communities are withheld — those are what you predict and submit.

The same data is also available in **long format**, which is often more convenient for modeling. In long format there is one row per community $\times$ species combination, with the columns `community`, `species`, `present` (0/1), plus the environmental predictors of that community repeated across its species:

```         
community  species  present  Bio1  Bio2  ...  Elevation
   212      sp_3       1      ...   ...        ...
   212      sp_7       0      ...   ...        ...
```

Wide and long are two views of the same data; you can always convert between them. For example, to turn a long label table back into the wide matrix:

```{r chunk_sdm_reshape, eval=FALSE}
load("sdm.RData")

library(missRanger)
library(tidyverse)

sdm_long |> summary()

sdm_long[,-3] = sdm_long[,-3]  |>  mutate(across(everything(), ~ifelse(is.na(.), mean(., na.rm = TRUE), .)))

train = sdm_long[!is.na(sdm_long$present), ]
test = sdm_long[is.na(sdm_long$present), ]

# sub sample training data due to computational constraints
indices = sample.int(nrow(train), 5000)
train$present = as.factor(train$present)

```



### A minimal working example

Dataset can be downloaded from the submission server or from [dropbox](https://www.dropbox.com/scl/fi/2juk5nk5tb87pzzfflqo3/sdm.RData?rlkey=oicnumxp93qmvnw85mpmunrf5&st=b1azbdlk&dl=0)

There are two natural modeling strategies:

1.  **Long format + a single binary classifier** with the species included as a predictor (e.g. a random forest or DNN on the long table). This gives a presence probability for every community $\times$ species row directly.
2.  **Wide format + a multi-output model** that predicts the full 2000 $\times$ 448 matrix at once.

A simple example using the long table:

```{r chunk_sdm_model, eval=FALSE}
library(ranger)
set.seed(123)

rf = ranger(present ~ ., data = train[indices, -c(1, 4)], # without community and survey ID
            classification = TRUE, probability = TRUE, num.trees = 100L) # only 100 trees to speed up the predictions of the test dataset

# presence probability per community x species row:
preds = predict(rf, data = test)$predictions[, 2]
```

### Submission format — long or wide


You can submit your predictions in **either** format; the leaderboard accepts both.


**Long** — the 896,000 (= 2000 $\times$ 448) probabilities stacked into a single column. The scorer rebuilds the matrix with `matrix(y, nrow = 2000)`, which fills it **column-by-column**, so the values must be ordered *all 2000 communities of species 1, then all of species 2, …* — exactly what `as.vector()` of the wide matrix produces:

```{r chunk_sdm_sub_long, eval=FALSE}
write.csv(data.frame(y = as.vector(preds)),
          file = "sdm.csv", row.names = FALSE)
```


**Wide** — a 2000 $\times$ 448 matrix of presence probabilities, communities in rows and species in columns, in the same species order as the label matrix (`sp_9`, `sp_53`, …):

```{r chunk_sdm_sub_wide, eval=FALSE}
write.csv(pred_matrix, file = "sdm.csv", row.names = FALSE)
```


Two checks before uploading: `dim(pred_matrix)` should be `2000 x 448`, and the saved file should have either 448 value columns (wide) or 896,000 rows (long). If you build the long vector directly from a long-format prediction (like the random forest above), make sure its rows are in community-within-species order, so the reshape lines up with the species columns of the label matrix — otherwise the values misalign and the score collapses.

**Evaluation metric:** submissions are scored with a **weighted AUC** — the mean of the per-species AUC values, weighted so that rare species count more than common ones.

Then submit the csv on <a href="http://132.199.73.15:8500" target="_blank" rel="noopener">http://132.199.73.15:8500</a>.

<!-- ## Wine -->

<!-- The data set is a collection of wines of different quality. The aim is to predict the quality of the wine based on physiochemical predictors. -->

<!-- For inspiration and data exploration notebooks, check out this <a href="https://www.kaggle.com/uciml/red-wine-quality-cortez-et-al-2009" target="_blank" rel="noopener">kaggle competition</a>. For instance, check out this very nice <a href="https://www.kaggle.com/aditimulye/red-wine-quality-assesment-starter-pack" target="_blank" rel="noopener">notebook</a> which removes a few problems from the data. -->

<!-- **Response variable:** "quality" -->

<!-- We could theoretically use a regression model for this task but we will stick with a classification model. -->

<!-- A minimal working example: -->

<!-- 1.  Load data set: -->

<!-- ```{r chunk_chapter8_14} -->
<!-- library(EcoData) -->

<!-- data(wine) -->
<!-- summary(wine) -->
<!-- ``` -->

<!-- 2.  Impute missing values (not our response variable!). -->

<!-- ```{r chunk_chapter8_15, message=FALSE, warning=FALSE} -->
<!-- library(missRanger) -->
<!-- library(dplyr) -->
<!-- set.seed(123) -->

<!-- wine_imputed = missRanger::missRanger(data = wine |> select(-quality), verbose = 0) -->
<!-- wine_imputed$quality = wine$quality -->
<!-- ``` -->

<!-- 3.  Split into training and test set: -->

<!-- ```{r chunk_chapter8_16} -->
<!-- train = wine_imputed[!is.na(wine$quality), ] -->
<!-- test = wine_imputed[is.na(wine$quality), ] -->
<!-- ``` -->

<!-- 4.  Train model: -->

<!-- ```{r chunk_chapter8_17, message=FALSE, warning=FALSE} -->
<!-- library(ranger) -->
<!-- set.seed(123) -->

<!-- rf = ranger(quality~., data = train, classification = TRUE) -->
<!-- ``` -->

<!-- 5.  Predictions: -->

<!-- ```{r chunk_chapter8_18} -->
<!-- preds = predict(rf, data = test)$predictions -->
<!-- head(preds) -->
<!-- ``` -->

<!-- 6.  Create submission csv: -->

<!-- ```{r chunk_chapter8_19, eval=FALSE} -->
<!-- write.csv(data.frame(y = preds), file = "rf.csv") -->
<!-- ``` -->

<!-- ## Nasa -->

<!-- A collection about asteroids and their characteristics from kaggle. The aim is to predict whether the asteroids are hazardous or not. For inspiration and data exploration notebooks, check out this <a href="https://www.kaggle.com/shrutimehta/nasa-asteroids-classification" target="_blank" rel="noopener">kaggle competition</a>. -->

<!-- **Response variable:** "Hazardous" -->

<!-- 1.  Load data set: -->

<!-- ```{r chunk_chapter8_20} -->
<!-- library(EcoData) -->

<!-- data(nasa) -->
<!-- summary(nasa) -->
<!-- ``` -->

<!-- 2.  Impute missing values (not our response variable!): -->

<!-- ```{r chunk_chapter8_21, message=FALSE, warning=FALSE} -->
<!-- library(missRanger) -->
<!-- library(dplyr) -->
<!-- set.seed(123) -->

<!-- nasa_imputed = missRanger::missRanger(data = nasa |> select(-Hazardous), -->
<!--                                       maxiter = 1, num.trees = 5L, verbose = 0) -->
<!-- nasa_imputed$Hazardous = nasa$Hazardous -->
<!-- ``` -->

<!-- 3.  Split into training and test set: -->

<!-- ```{r chunk_chapter8_22} -->
<!-- train = nasa_imputed[!is.na(nasa$Hazardous), ] -->
<!-- test = nasa_imputed[is.na(nasa$Hazardous), ] -->
<!-- ``` -->

<!-- 4.  Train model: -->

<!-- ```{r chunk_chapter8_23, message=FALSE, warning=FALSE} -->
<!-- library(ranger) -->
<!-- set.seed(123) -->

<!-- rf = ranger(Hazardous~., data = train, classification = TRUE, -->
<!--             probability = TRUE) -->
<!-- ``` -->

<!-- 5.  Predictions: -->

<!-- ```{r chunk_chapter8_24} -->
<!-- preds = predict(rf, data = test)$predictions[,2] -->
<!-- head(preds) -->
<!-- ``` -->

<!-- 6.  Create submission csv: -->

<!-- ```{r chunk_chapter8_25, eval=FALSE} -->
<!-- write.csv(data.frame(y = preds), file = "rf.csv") -->
<!-- ``` -->

## Flower

A collection of over 4000 flower images of 5 plant species. The data set is from <a href="https://www.kaggle.com/alxmamaev/flowers-recognition" target="_blank" rel="noopener">kaggle</a> but we downsampled the images from $320*240$ to $80*80$ pixels. You can a) download the data set <a href="http://132.199.73.15:8500" target="_blank" rel="noopener">here</a> or b) get it via the EcoData package.

**Notes:**

- Check out convolutional neural network notebooks on kaggle (they are often written in Python but you can still copy the architectures), e.g. <a href="https://www.kaggle.com/alirazaaliqadri/flower-recognition-tensorflow-keras-sequential" target="_blank" rel="noopener">this one</a>.
- Last year's winners have used a transfer learning approach (they achieved around 70% accuracy), check out this <a href="https://www.kaggle.com/stpeteishii/flower-name-classify-densenet201" target="_blank" rel="noopener">notebook</a>, see also the section about transfer learning \@ref(transfer).

**Response variable:** "Plant species"

1.  Load data set:

```{r chunk_chapter8_26, message=FALSE, warning=FALSE}
library(tensorflow)
library(keras3)

train = EcoData::dataset_flower()$train/255
test = EcoData::dataset_flower()$test/255
labels = EcoData::dataset_flower()$labels
```

Let's visualize a flower:

```{r chunk_chapter8_27}
train[100,,,] |>
  image_to_array() |>
  as.raster() |>
  plot()
```

2.  Build and train model:

```{r chunk_chapter8_28, eval=FALSE, warning=FALSE}
model = keras_model_sequential()
model |> 
  layer_conv_2d(filters = 4L, kernel_size = 2L,
                input_shape = list(80L, 80L, 3L)) |> 
  layer_max_pooling_2d() |> 
  layer_flatten() |> 
  layer_dense(units = 5L, activation = "softmax")

### Model fitting ###

model |> 
  compile(loss = loss_categorical_crossentropy, 
          optimizer = optimizer_adamax(learning_rate = 0.01))

model |> 
  fit(x = train, y = keras::k_one_hot(labels, 5L))


```

3.  Predictions:

```{r chunk_chapter8_29, eval=FALSE}
# Prediction on training data:
pred = apply(model |> predict(train), 1, which.max)
Metrics::accuracy(pred - 1L, labels)
table(pred)

# Prediction for the submission server:
pred = model |> predict(test) |> apply(1, which.max) - 1L
table(pred)
```

4.  Create submission csv:

```{r chunk_chapter8_30, eval=FALSE}
write.csv(data.frame(y = pred), file = "cnn.csv")
```