Tech-enabled Personalized Learning in Low- and Middle-income Countries – A UNICEF Landscape Analysis

• In Event: Trends in the ICT Field

Proposal

I. Theory/context

The idea of quality and equitable education has been a long-standing goal for educators and is central to the vision of SDG4. However, millions of children are out of school or falling farther behind in their learning due to the factors associated with the pandemic (Vegas, 2020). The 2022 CIES theme of idea/lism encourages a shifting in thought that embraces innovation and creativity in settings that have previously relied on compliance and conformity. In many parts of the world, formal schooling has been delivered with relatively little modification since its inception (Winthrop & McGivney, 2015). This moment lends well to a re-envisioned approach to learning that allows students to engage in more personalized and adaptive educational processes.

This study leverages research from earlier work on tech-enabled personalized learning systems (Bulger, 2016) to interrogate how personalized learning design can be deployed in low-resourced educational settings. Although definitions of personalized learning (PL) vary, there is significant agreement that it is learner-centered and flexible, and responsive to individual needs. While PL on some level has existed in pockets for quite some time, advancements in ICT, data science and machine learning have led to a significant growth in the development of tech-enabled solutions to support personalized learning in and out of schools (Kharas & Cohen, 2018). While evidence is emerging about the use of technology to personalize learning, much of the literature on PL is centered on high-income countries in well-resourced educational settings.

This research presents a landscape analysis of emerging trends, challenges and opportunities to effective design, implementation and scaling of PL solutions specific to low- and middle-income countries (LMICs). The discussion of findings for this paper emphasizes approaches to adapting learning paths through emerging technologies including machine learning, artificial intelligence, and natural language processing. This research seeks to provide balanced insights that will inform policy discourse and help shape the landscape on successful PL design and implementation in LIMCs.

II. Inquiry

A. Instrument design

Data is drawn from a global landscape review of forty PL platforms that were reviewed with a project-specific mapping instrument. The instrument was developed through a literature review that identified relevant elements of personalized learning. Additionally, fifteen ed-tech frameworks were compared to analyze the classification of distinct learning and technology features. This analysis produced a unique mapping rubric of over 140 elements that combined features of the previous frameworks into an original instrument relevant to the present study objectives. Rubric items are distributed along eleven categories encompassing features related to design, use, and inclusivity.

B. Sampling

Inclusion parameters for the landscape review followed a criteria sampling approach that limited the scope of the research to primary and secondary school settings in LMIC contexts, and that were beyond the pilot implementation phase. Initially, we searched through a series of conference websites and website aggregators to identify projects that claimed to apply some aspect of tech-enabled personalized learning that produced a list of 1,000 potential projects. From there, we applied our inclusion parameters to reduce the sample to about 150. The project team then completed a more detailed review of the reduced sample based on relevance to the research objectives and equitable distribution by region. A final sample of forty PL projects were included in this study.

C. Analysis

A team of mappers collected data on each of the forty PL projects along the 140 item rubric. Data was then analyzed along twelve themes relevant to the research objectives. Given the nature of the non-representative sampling approach, this study did not intend to derive inferences about PL effectiveness from the data. Instead, cross-tabulations were used to classify and categorize emerging trends from the research. While the full report details each of the twelve themes from the study, this paper will discuss findings specific to the approaches used to promote tailored learning through emerging technologies.

III. Findings

For the adaptation theme, our research produced insights across two approaches: AI-enabled adaptation and limited algorithm-enabled adaptation.

AI-enabled adaptation is becoming more common in developing countries. About 70% of the products reviewed in our study utilized AI in their adaptation approach. This approach involves tools that leverage either Machine Learning (74%), Natural Language Processing (around 15%) or Expert Systems (7%) to monitor and assess students' learning, and adapt the learning pathway based on AI-enabled predictive analytics (75%). This approach also includes AI-driven data processing methods to provide feedback and point students to resources to reinforce skills (around 63%). For instance three products, Joven Gênios, Siyavula and Gooru, use AI for knowledge tracing to identify the learner's inferred proficiency level and generate the most optimal learning path. Some of the algorithms used to perform these tasks include Support Vector Machine (SVM), Artificial Neural Networks (ANN), Recurrent Neural Networks (RNNs) for deep knowledge tracing, and Naive Bayes. Other uses for AI cited in our study include text to speech algorithms to assess learners who have difficulties with reading.

The limited algorithm-enabled adaptation on the other hand is less widely used in developing countries, but is well suited for adaptation. Limited adaptation approach includes solutions that offer adaptation through “if-then” rules. Around 22% of the products follow a limited algorithm approach where static learning paths are set at intervals throughout the learning journey, thus allowing learners to advance or regress individually.

The study also revealed common challenges faced by AI technologies in LMIC contexts. One significant challenge to accelerating the use of AI-enabled adaptation relates to the quality and quantity of data to train the algorithms appropriate to the local context. Lack of large, comprehensive and representative training data not only inhibits the accuracy of the predictions, but it also creates opportunities for bias within these platforms. Furthermore, our findings reveal a large variation in perspective on what standards AI technologies should follow for deployment across LMICs. Other PL challenges and their implications for deployment in low-resourced settings are discussed.

Authors

• Nathan M. Castillo, University of Illinois

• Ghaida Alrawashdeh, University of Illinois Urbana-Champ