User Profiling in Social Media
Overview
This project was completed as part of TCSS555 at the University of Washington, Tacoma, in collaboration with Sammy Franklin and Lang Wang. The goal was to build a system that infers user characteristics — age range, gender, and the Big Five personality traits (OCEAN) — from Facebook data provided by the myPersonality dataset. Each team member owned a different data modality: I handled Facebook Likes (relation data), Sammy handled status text, and Lang handled profile pictures.
My Contribution: Inference from Liked Links
My responsibility was to develop the full pipeline for predicting age range and gender from users’ Facebook liked links (relation data).
Data Processing
The raw relation data mapped user IDs to liked page links. I restructured this into a user-centric format using Pandas: iterating through the relation data to consolidate each user’s likes into a single row, then merging with the profile data via an inner join. For age inference, I applied a piecewise bucketing function to convert raw ages into four groups (under 25, 25–34, 35–49, 50+), encoded as integer labels for compatibility with sklearn models.
Model Training
The consolidated like-link strings were fed into a CountVectorizer, producing a sparse binary matrix (users × unique like IDs) used as the feature matrix X_train. I experimented with several classifiers:
- Naive Bayes (initial baseline)
- Logistic Regression
- XGBoost
- Random Forest
- Support Vector Machines (gender only)
Ultimately, an ensemble using majority voting across Naive Bayes, Logistic Regression, and XGBoost yielded the best results. Both the vectorizer and trained models were serialized with Pickle for use at inference time.
Prediction & Ensemble Voting
At prediction time, each model in the ensemble produced an independent prediction, stored in a DataFrame. The mode across models was taken as the final prediction — a simple but effective voting strategy. Age predictions were converted back to their string group labels before submission.
Experiment: Removing Infrequent Likes
I ran an ablation experiment to test whether removing the least-frequently liked links improved accuracy. Removing up to 5 of the rarest links showed a modest improvement in gender accuracy locally, but hurt age range accuracy — and both metrics declined on the evaluation VM, suggesting the removal introduced noise rather than reducing it.
Results
| Test Type | Age Range Accuracy | Gender Accuracy |
|---|---|---|
| Training Data | 0.63 | 0.82 |
| Evaluation Data | 0.66 | 0.82 |
Gender, as a binary classification task, was the stronger result at 0.82 on both training and evaluation data. Age range accuracy improved by 3% from training to evaluation, finishing at 0.66 across four categories.
Team Results Summary
| Modality | Model | Age Acc | Gender Acc | OCEAN Avg RMSE |
|---|---|---|---|---|
| Likes (Dean) | Ensemble (NB + LR + XGBoost) | 0.66 | 0.82 | — |
| Text (Sammy) | MLP-2 (tag-partitioned sums) | 0.562 | 0.731 | 0.726 |
| Text (Sammy) | LSTM + LIWC | 0.508 | 0.675 | 0.716 |
| Images (Lang) | 2D CNN (Keras/TensorFlow) | — | 0.65 | — |
The like-based ensemble achieved the highest gender accuracy across all modalities, and every team member exceeded the baseline for at least one category.
Tech Stack
- Python, Pandas, scikit-learn, XGBoost
- PyTorch, NLTK, GloVe embeddings (text models)
- Keras / TensorFlow (CNN for images)
- Pickle for model serialization
- Ubuntu server for evaluation
