This project develops an intelligent Targeted Advertising System designed to enhance user engagement and conversion rates by delivering highly relevant ad content. The system achieves this by analyzing user behavior, semantically clustering users into distinct groups, and then predicting the preference level of each user segment for various advertising banners. The foundational principle is that users exhibiting similar behavioral patterns within an application are likely to share similar advertising interests.
-
Semantic User Clustering: Groups users into meaningful segments based on their in-app behavior.
-
Banner Content Embedding: Transforms raw textual ad banner captions into numerical vector representations.
-
Preference Prediction Model: A regression model that predicts the likelihood of a user segment interacting positively with a given ad banner.
Users are categorized into 500 distinct semantic clusters based on their behavioral data within the application. Each user is initially represented as an 8-dimensional vector in a semantic space. For each of these 500 clusters, a representative segment vector (s_i) is computed (e.g., by averaging the vectors of all users belonging to that cluster). This approach allows for a high-level, generalized understanding of user preferences.
The textual captions of advertising banners are converted into numerical vector embeddings (b_j). This process leverages Natural Language Processing (NLP) techniques, such as TF-IDF or more advanced pre-trained language models, to capture the semantic meaning of the ad content. The goal is to represent each banner in a numerical space where similar banners are located closer to each other.
A regression model, denoted as f, is designed to predict the preference score ( hatp_ij) for a given user segment i and banner j. The model takes the user segment's representative vector (s_i) and the banner's embedding vector (b_j) as inputs.
The model is trained on a provided dataset (p_train) containing known (segment, banner) pairs and their corresponding preference values. The objective is to learn the underlying patterns that dictate user segment preferences for different ad creatives.
Training and Evaluation The model is trained using the p_train dataset. Its performance is evaluated by predicting preference scores for unseen (segment, banner) pairs in the p_test dataset. Standard regression evaluation metrics such as Mean Absolute Error (MAE) and R-squared can be used to assess the model's accuracy and predictive power.
For potential improvements in model performance, exploring architectures inspired by Neural Collaborative Filtering (NCF) is recommended. Additionally, experimenting with various advanced NLP techniques for banner caption embeddings (e.g., transformer-based models) can yield better semantic representations.
This project's dataset includes:
-
User Matrix: Over 300,000 users, each represented in an 8-dimensional semantic space.
-
Banner Captions: Persian text captions for all advertising banners.
-
User Segment Array: An index mapping each user to one of the 500 predefined semantic clusters.
-
p_train DataFrame: Training data consisting of (segment, banner) pairs with their associated preference scores.
-
p_test DataFrame: Test data containing (segment, banner) pairs for which preference scores need to be predicted.
The program will output a CSV file containing the predicted preference scores for the p_test pairs.
NOTE: The estimated error for the output CSV file was 0.138, indicating strong model performance.