--- title: "Introduction to mccca" output: rmarkdown::html_vignette vignette: > %\VignetteIndexEntry{Introduction to mccca} %\VignetteEngine{knitr::rmarkdown} %\VignetteEncoding{UTF-8} --- ```{r, include = FALSE} knitr::opts_chunk$set( collapse = TRUE, comment = "#>", fig.width = 7, fig.height = 6, out.width = "90%" ) ``` Multiple-class cluster correspondence analysis (MCCCA: Takagishi and van de Velden, 2022) is a statistical method that identifies and visualizes heterogeneous tendencies specific to certain "classes", specified by the researcher, in a low dimensional space. In MCCCA, two kinds of variables, external and active variables, are distinguished. External variables directly define the classes whereas the active variables are used to derive, for each class, multiple clusters of observations exhibiting similar tendencies. The class-specific clusters are displayed together with the categories of the active variables in a single biplot. In order to apply MCCCA, categorical data are required. Moreover, the data should consist of two types of variables: active and external variables. The external variables, and interactions between them, define the known classes. The total number of, and the distribution over, these classes may not always be obvious. Hence, the general flow of function usage is designed to first transform the data and inspect these dimensions. In particular, first, the data set needs to be transformed to an `mcccadata` object using `create.MCCCAdata()`. Next, `MCCCA()` is applied to the `mcccadata` object after which the results of `MCCCA()` can be visualized using `plot()`. A numerical summary can be obtained using `summary()`. To specify an appropriate number of clusters for each class, a dedicated tuning function `decideK()` can be used. | Function | Description | |---|---| | `create.MCCCAdata()` | Creates an `mcccadata` object which is later passed to `MCCCA()`. | | `MCCCA()` | Applies MCCCA to the `mcccadata` object created by `create.MCCCAdata()`. | | `decideK()` | Calculates the cluster index (KL index by default) to determine the number of class-specific clusters. | | `summary()` | Produces a detailed output for `create.MCCCAdata()`, `MCCCA()` and `decideK()`. | | `plot()` | Shows a biplot using quantifications estimated by `MCCCA()`. | ## Example: Meal and drink preferences The `mealDrink` data, described in Section 3 of Takagishi and van de Velden (2022) and available in the package, consists of artificial data for 390 observations on 4 categorical variables. For these data, the aim is to identify whether there are different tendencies in meal and drink preferences (the active variables) depending on nationality and gender (the external variables). ```{r} library(mccca) data(mealDrink) ``` | Variable type | Variable name | Category | |---|---|---| | Active variables | Meal | Meat, Fish, Rice | | | Drink | Fruit juice, Tea, Milk | | External variables | Nationality | American, Japanese | | | Gender | Female, Male | As we have two binary external variables, four classes can be made. These classes correspond to American females, American males, Japanese females, and Japanese males. The characteristics of these data are as follows: most Americans prefer Meat and Juice, while a smaller group of American females select Fish and Tea. Similarly, the majority of Japanese prefer Rice and Tea, while a small group of the Japanese females chooses Rice and Milk. Instead of only focusing on majority preferences, MCCCA allows us to visualize heterogeneous tendencies for the subgroups defined by the categories of the external variables. Note that the `mealDrink` data set in mccca consists of four categorical variables. The first two variables are active and the last two are external. To apply `MCCCA()`, separate data frames consisting of the active and external variables are required. We can create these as follows: ```{r} active <- c(1, 2) external <- c(3, 4) dat.act <- mealDrink[, active] dat.ext <- mealDrink[, external] ``` The first rows of these data are: ```{r} head(dat.act) head(dat.ext) ``` Note that, when merging active and external variables from separate sources, the rows of the two data frames should correspond. To apply MCCCA, we need to construct the classes as defined by the external variables. Rather than having this done internally in the `MCCCA()` function, this needs to be done by applying the `create.MCCCAdata()` function. In that way, before applying MCCCA, the user is presented with an overview of the constructed classes as well as a distribution over these classes. ```{r} mcccadata <- create.MCCCAdata(dat.act, dat.ext) ``` The data corresponding to the active and external variables are required arguments in `create.MCCCAdata()`. By applying `create.MCCCAdata()`, all classes (corresponding to the combinations of categories of the external variables) are created, and the numbers of observations belonging to each class are printed. Furthermore, the function results in an object, here `mcccadata`, which is a list consisting of the data set split by class, and the information for each class. An overview of the resulting object can be obtained using the `summary()` function: ```{r} summary(mcccadata) ``` To apply MCCCA, the number of class-specific clusters needs to be specified using an integer vector of the appropriate length. The total number of "external" classes, consisting of all combinations of the external classes, is stored as `mcccadata$C` and can also be found in the output of `summary()`. For this example, the number of classes is C = 4, and we therefore need to define an integer vector of length 4 containing the number of class-specific clusters that we aim to retrieve. The numbers of the class-specific clusters need to be supplied by the user. As in most cluster methods, this choice is not trivial. For the `mealDrink` data, however, we can use the "true" (i.e., the numbers used to generate the data) numbers for the clusters. These true numbers are: 2 (American females), 1 (American males), 2 (Japanese females) and 2 (Japanese males). After specifying this as `K.vec`, we perform MCCCA: ```{r} K.vec <- c(2, 1, 2, 2) set.seed(1) res <- MCCCA(mcccadata, K.vec = K.vec) ``` The `MCCCA()` function requires an `mcccadata` object (i.e., a list produced by `create.MCCCAdata()`) and a vector with class-specific numbers of clusters `K.vec`. As the algorithm starts from random initial cluster assignments, we used `set.seed()` to make the results reproducible. We can use `plot()` to visualize the results of the MCCCA analysis. The resulting biplot includes labels for the class-specific clusters. By default, these labels are formed by combining the categories of the external variables (e.g., "American&female"). However, the labels may be customized using the `classlabel` argument. For our example, we can create more compact labels by only using the first letter of each class: ```{r, fig.alt = "MCCCA biplot of the mealDrink data showing category points for meal and drink and bubbles for the seven class-specific clusters."} shortlabels <- c("AF", "AM", "JF", "JM") plot(res, classlabel = shortlabels) ``` Different colors and symbols are used to distinguish between categories of the active variables and class-specific clusters. In particular, triangle points correspond to categories of the active variables, and categories belonging to different variables are displayed using different colors. For the class-specific cluster points, colors indicate the classes. Moreover, each class-specific cluster center is represented by a bubble. The bubble sizes are proportional to the cluster sizes, and cluster numbers are assigned according to relative cluster size, where a value of 1 indicates the largest cluster within a class. A legend showing the class-color correspondence and a scale for the bubble sizes is added automatically. Note that the order of `shortlabels` should be the same as the order of `classlabel` in the `mcccadata` object (i.e., `mcccadata$classlabel`). There are many other options for customization of the visualization using the `plot()` function. See `?plot.mccca` for a complete overview. To inspect numerical results, in particular coordinates and fit per dimension, one can use `summary()`. ```{r} summary(res) ``` The output of `summary()` consists of the estimated coordinates of class-specific cluster centroids and active variable category points, i.e., `Gg` and `Bg`, respectively. Note that the "gamma-scaled" coordinates are scaled in such a way that the average squared deviation from the origin of the row and column points is the same (see Section 2.3 in Takagishi and van de Velden, 2022, for details). Furthermore, "American&female1" indicates the largest cluster in the "American and female" class. In the table with cluster centroid coordinates, cluster sizes and relative cluster sizes within each class are also listed. Here, we see, for example, that 10 observations belong to the smaller cluster of the "Japanese female" class. To assess the quality of the two-dimensional visualization, the relative percentages of adjusted inertias, as proposed by Benzécri (1979) (also described in Greenacre (1984)), can be shown along the dimensions of the biplot by `plot(res, show.inertia = TRUE)`. More detailed information on the distribution of inertia is stored in the `inertia` field. These adjusted inertias are obtained by applying MCA to the full data set. For our example, we have: ```{r} res$inertia ``` Here, the first column gives the adjusted, the second the relative, and the third the cumulative inertia. ### The averaging approach To further illustrate the use of MCCCA and how visualizing the class-specific heterogeneity improves interpretation, we apply the averaging approach and compare results. As the averaging approach corresponds to MCCCA with the numbers of all class-specific clusters set to 1, we can achieve this by executing: ```{r} K.vec <- c(1, 1, 1, 1) res.ave <- MCCCA(mcccadata, K.vec = K.vec) ``` We can plot the results again by using `plot()`. ```{r, fig.alt = "Biplot of the averaging approach for the mealDrink data with one bubble per class."} plot(res.ave, classlabel = shortlabels, main = "Averaging result") ``` Comparing the averaging approach with the MCCCA results shows that the averaging approach only reveals a single association pattern for each class, whereas MCCCA identifies heterogeneous tendencies within classes. For example, the averaging approach associates the Japanese female class with "Rice" and "Tea". In contrast, MCCCA reveals two distinct clusters within this class: a larger cluster associated with "Rice" and "Tea" and a smaller cluster associated with "Rice" and "Milk". Similarly, for American females, the two clusters differ with respect to meal preferences (i.e., Fish or Meat). ## Choosing the numbers of clusters In the example above, we used the "true" number of clusters. In most real data applications, the number of clusters is unknown. Determining an appropriate number is, as in any clustering method, a difficult task as there is usually no objective validation measure available. The mccca package provides the function `decideK()` to assist in determining the number of clusters. In particular, `decideK()` performs k-means clustering on MCA results (i.e., the row coordinates) based on observations for each class, using the active variables, for a specified range of values for the number of clusters. It calculates the value for a specified cluster quality index for each number of clusters in the range, and uses these to suggest a choice for the number of clusters. See `?decideK` for details and an example. ## References Benzécri, J.-P. (1979). Sur le calcul des taux d'inertie dans l'analyse d'un questionnaire. *Cahiers de l'Analyse des Données*, 4, 377--378. Greenacre, M. J. (1984). *Theory and Applications of Correspondence Analysis*. Academic Press, London. Takagishi, M. and van de Velden, M. (2022). Visualizing class specific heterogeneous tendencies in categorical data. *Journal of Computational and Graphical Statistics*, 31(3), 790--801.