Smart Recommendation System Based on Understanding User Behaviour for Afan Oromo Language with Deep Learning

Why a Dedicated System for Afan Oromo?

Most mainstream recommendation platforms are built around highresource languages such as English, Mandarin, or Spanish. Consequently, they suffer from two major shortcomings when applied to Afan Oromo:

• Data scarcity: Publicly available corpora, clickstream logs, and collaborativefiltering datasets are limited.
• Cultural mismatch: Recommendation logic that ignores local customs, dialectal variations, and content relevance can produce irrelevant or even offensive results.

Addressing these gaps requires a system that can:

1. Learn from sparse, noisy interaction data.
2. Incorporate linguistic features specific to Afan Oromo (e.g., verb morphology, honorifics).
3. Adapt to evolving user interests while preserving privacy.

Core Architecture Overview

The proposed recommendation pipeline consists of four interconnected modules:

1. Data Collection & Preprocessing

• Interaction logs: clicks, watch time, rating, dwell time on Afan Oromo articles, videos, podcasts, and elearning modules.
• Explicit feedback: thumbsup/down, star ratings, and short textual reviews.
• Contextual signals: time of day, device type, location (region within Ethiopia/Kenya), and language variant (e.g., WestCentral vs. Eastern dialects).

Preprocessing steps include tokenisation with a languagespecific tokenizer, removal of stopwords, and conversion of textual feedback into sentiment scores using a finetuned BERT model for Afan Oromo.

2. User Behaviour Modelling

Two complementary deeplearning models are employed:

• Sequential Neural Networks (e.g., GRU, Transformer): Capture the order of user actions, detecting patterns such as watchthensearch or readthenshare.
• Graph Neural Networks (GNN): Represent users and items as nodes in a bipartite graph, allowing the system to propagate preferences through shared neighbours (e.g., users who liked the same cartoon series).

3. Content Representation

Deep language models trained on Afan Oromo corpora provide dense embeddings for every item:

• OromoBERT a BERTstyle encoder trained on news, literature, and socialmedia text.
• Multimodal encoders for video (audiotranscript + visual features) and audioonly podcasts.

These embeddings enable semantic similarity matching beyond simple keyword overlap, crucial for languages with rich morphology.

4. Ranking & Recommendation Engine

The final score for an item i and a user u is a weighted combination of:

• Behavioural affinity f_behaviour(u,i) from the sequential/GNN models.
• Content similarity f_content(u,i) based on embedding cosine similarity.
• Contextual adjustment f_context(u,i) (timeofday, dialect match).

Training uses a pairwise ranking loss (e.g., Bayesian Personalized Ranking) together with regularisation that penalises overexposure of any single content source.

Addressing Data Scarcity

Several strategies mitigate the lack of largescale Afan Oromo datasets:

1. Transfer learning: Initialise the language encoder with weights from a multilingual model (e.g., XLMR) and finetune on the limited Oromo corpus.
2. Data augmentation: Backtranslation, synonym replacement, and morphological inflection generation increase training samples without altering meaning.
3. Crossdomain bootstrapping: Leverage user behaviour on bilingual platforms (e.g., OromoEnglish news sites) to infer preferences for pure Oromo content.

PrivacyPreserving Personalisation

Respecting user privacy is essential, especially in regions where digital rights are still developing. The system incorporates:

• Federated Learning: Model updates are computed locally on the users device and aggregated serverside, preventing raw interaction logs from leaving the phone.
• Differential Privacy: Gaussian noise is added to gradient updates, providing mathematical guarantees that individual actions cannot be reverseengineered.
• Ondevice caching: Frequently accessed embeddings are stored locally, reducing network traffic and latency.

Evaluation Metrics

Beyond standard accuracy metrics (Hit Rate, NDCG), the following measures capture the unique goals of an Oromofocused system:

• Cultural Relevance Score (CRS): Human judges rate recommendations on a 5point scale for cultural appropriateness.
• Dialect Coverage (DC): Proportion of recommended items that match the users preferred dialect.
• Content Diversity Index (CDI): Entropybased metric ensuring the system does not overrecommend a single genre (e.g., only news).

Offline tests on a heldout dataset are complemented by online A/B tests measuring clickthrough rate (CTR), session length, and churn reduction.

Potential Applications

1. Elearning portals: Suggest lessons, storybooks, and interactive exercises aligned with a learners progress and dialect.

2. Streaming services: Recommend Oromo movies, music videos, and podcasts while surfacing emerging local creators.

3. News aggregators: Personalise news feeds, highlighting communityrelevant stories (e.g., agricultural tips, health alerts).

4. Ecommerce: Suggest products with descriptions in Afan Oromo, increasing trust and conversion among native speakers.

Challenges and Future Directions

Coldstart problem. New users and items lack interaction history. Hybrid approaches that combine contentbased similarity with demographic cues (age, region) can provide initial recommendations.

Dialectal variation. Continuous collection of dialectspecific corpora and dynamic updating of embeddings will improve matching accuracy.

Explainability. Users should understand why an item is suggested. Attentionvisualisation tools from the Transformer model can highlight which past actions influenced the current recommendation.

Future research may explore:

• Reinforcement learning agents that adapt to longterm user satisfaction.
• Multilingual joint training with Amharic and Somali to share linguistic structures.
• Communitydriven curation where local content creators label items for relevance, feeding back into the model.

Conclusion

Building a smart recommendation system for Afan Oromo demands a blend of deeplearning techniques, cultural awareness, and privacyfirst engineering. By modelling sequential behaviour, leveraging graphbased relationships, and employing languagespecific embeddings, the proposed architecture can deliver highly relevant, diverse, and respectful suggestions across education, entertainment, and commerce domains. The ultimate impact is twofold: enriching the digital lives of Oromo speakers and fostering a vibrant ecosystem of locally produced content.