Migration in Mozambique

CSCI 5523 Final Project (University of Minnesota, Fall 2025)

Gabe Buchmann, Drew Gjerstad, Arlan Hegenbarth, Aaron Meyerhofer, and Hasan Raza

Important

The data used in this project is sourced from IPUMS International. Access to their data repositories is restricted according to their usage license. To gain access, apply for an account using this form. Once you are registered, you will need your API key to generate and download a data extract using the source code in this repository. You can obtain your API key from this site.

Introduction

In this project, we focus on the prediction (more generally, the estimation) of individuals who have recently migrated in Mozambique based on data collected by the Instituto Nacional de Estatística (INE) in 1997, 2007, and 2017 which was made available by IPUMS International (referred to herein as the Mozambique dataset). Census microdata, such as this dataset, is unique since it is made up of features collected at the individual level, rather than aggregated statistics that are typical in other census datasets. Furthermore, the 2017 sample in this dataset was only recently released meaning that the relationships are relatively unexplored.

We are primarily focusing on the exploration of methods used to extract insights and inferences from a large, high-dimensional, and temporal dataset. The Mozambique dataset should be great for this as it spans three decades and has numerous potential features we can use in our analyses. Furthermore, we are also aiming to develop a principled data preprocessing pipeline for the dataset as there are several considerations that must be made.

Our Approaches

The diagram below provides a high level overview of our approach to analyzing the Mozambique dataset using classification analysis. For our association analysis, we use the same data preprocessing shown on the left side of the figure and then follow standard procedures to generate candidate itemsets which are pruned using the Apriori algorithm for our association analysis. These steps are outlined in the association_analysis.ipynb notebook.

Figure 1: Data Preprocessing Pipeline

Data Preparation

In this project, we used two main approaches to analyze the Mozambique dataset: classification analysis and association analysis. However, before applying any analytical techniques, we needed to the preprocess the dataset as the raw data is incompatible with many of these techniques. First, we obtain the the raw data extract from IPUMS International. Then, we removed rows missing response (MIG1 and MIG5) values and any metadata or "detailed" columns which are not useful predictors. Note that "detailed" columns are more detailed versions of other columns but are not usually consistent across the census years and are thus not practical for our analyses. Next, we performed feature selection that was primarily motivated by several features having more than 90% missing values.

After feature selection we formatted the features to handle the various attribute types in our dataset including binary, categorical, and continuous. Some of our features were already in some binary form (i.e., Yes or No, etc.) but needed to be converted to integer zeros and ones for consistency in our analyses. For categorical variables, we used the attribute value-pair approach to convert the categories into new features. Finally, we used binning methods to bin continuous features in our dataset. With the features sufficiently processed for our analysis methods, we saved the dataset into locally stored Python serial (pickle, .pkl) files for development purposes.

Classification Analysis

For our classification analysis we tried three different models: a random forest model, a support vector classifier, and an artificial neural network model. Note that during our prototyping and development phase, we had trouble getting the support vector classifier to converge on any sort of solution due to the nature of our dataset. Therefore, we dropped that model from our analysis which is why it is not shown in the diagram above.

Random Forest

For our Random Forest, we used the RandomForestClassifier implementation from Scikit-Learn and performed hyperparameter tuning using grid search cross-validation on the following hyperparameters: number of estimators (trees) in the forest, maximum depth of each tree, minimum number of samples per split, and minimum samples per leaf node. We also used balanced class weights and bootstrapping. See msi_run.txt for output from the random forest models on each dataset.

Neural Network

For our neural network model, we use a simple architecture shown in the figure below: four fully-connected layers (two of which are hidden layers) connected by a sequence of batch normalization, a ReLU activation of each layer's output, and dropout with $p=0.3$ to prevent model overfitting. We train our network across 20 epochs with batches of 1024 samples and a learning rate of $\ell=0.001$. We use cross-entropy loss with class weightings to account for our severely imbalanced dataset and the Adam optimizer for optimizing our network weights. Finally, we incorporate a learning rate scheduler that reduces the learning rate as the network "stops learning" and incorporates "patience" for early stopping.

Figure 2: Neural Network Architecture

Association Analysis

For our association analysis, we use the same data preprocessing workflow shown on the left-hand side of Figure 1. Similar to the classification analysis, each dataset was split into training, validation, and test sets. The goal of our association analysis was to identify rules that were not general associations as is typical in standard association analysis. Rather, we focus on finding rules of this form: "Given these attributes and their corresponding values, the respondent has migrated." To begin, we run the Apriori algorithm on the training set. However, since our data has millions of records and 65 attributes, the traditional Apriori algorithm was not computationally tractable. Instead, we reformulate the problem given that we are focusing on rules that imply whether or not a respondent has migrated within the last year or within the last five years (MIG1 or MIG5, respectively). That is, we filter the dataset to include only rows corresponding to respondents that migrated and apply the Apriori algorithm to those records.

We applied the Apriori algorithm using a minimum support threshold of 25%. This translates to identifying a subset of attributes in each dataset such that at least 25% of respondents that migrated had this subset. Note that while this threshold is higher than typical thresholds for association analysis, our goal of finding broadly generalizable rules motivates a higher minimum threshold---in addition to the fact that we have already filtered the dataset. The 6,000 candidate rules produced for each dataset were evaluated using metrics including support, lift, confidence, and interest metrics. Furthermore, we also consider the length of the rule (i.e., how many attributes are in the rule's antecedent) since shorter rules are much more interpretable in this application. Finally, we applied each candidate rule as a de facto classifier on the validation set to find the "best" rule to use as a classifier on the unseen data in the test set.

Results

	Accuracy	Precision	Recall	F1 Score	AUC Score
Random Forest	0.7901	0.9830	0.7901	0.8723	0.7818
Neural Network	0.7674	0.9841	0.7674	0.8578	0.8108
Association Rule	0.8414	0.5071	0.7345	0.4590	0.5792

Table 1: Test Set Performance for Each Model Predicting MIG1

	Accuracy	Precision	Recall	F1 Score	AUC Score
Random Forest	0.7221	0.9436	0.7221	0.8029	0.8068
Neural Network	0.7093	0.9451	0.7093	0.7938	0.8139
Association Rule	0.7964	0.5112	0.6641	0.4941	0.5377

Table 2: Test Set Performance for Each Model Predicting MIG5

Figure 3: Test Set ROC Curves

Figure 4: Test Set Confusion Matrices