http://hdl.handle.net/123456789/13169 Sat, 04 Apr 2026 14:11:30 GMT 2026-04-04T14:11:30Z http://hdl.handle.net/123456789/19845 Recommendation System Using Personalized Chat-Bot Muddasir Aziz Khattak, 01-249231-007 Predicting a person’s personality can significantly enhance personalized recommendations across various domains. This study introduces PersonalityBERT, a deep learning model designed to determine personality traits from user-generated text. We utilized the MBTI (Myers-Briggs Type Indicator) dataset, which categorizes personalities into 16 distinct types, and fine-tuned BERT (Bidirectional Encoder Representations from Transformers) to analyze text input effectively. BERT’s contextaware and transformer-based multi-head self-attention mechanisms enable a deeper understanding of semantic meaning, improving personality classification accuracy.To extract meaningful personality insights, preprocessing techniques such as stopword removal, lemmatization, and text vectorization were applied. The fine-tuned BERT model was trained on the MBTI dataset and optimized using hyperparameter tuning, including learning rate adjustments, gradient accumulation, and weight decay regularization. Our PersonalityBERT model achieved an accuracy of 88%, surpassing baseline models, including traditional NLP classifiers and LSTMs. Additionally, the model produced a precision of 0.86, recall of 0.90, F1-score of 0.88, and an AUC-ROC score of 0.94, demonstrating high reliability in personality classification. Beyond classification, this model was integrated into a personality-based recommendation system. Using Sentence Transformers and Cosine Similarity, PersonalityBERT provides personalized book recommendations based on a user’s personality type and textual input. The chatbot framework was developed using Streamlit, with backend deployment. The model successfully matches users to book reviews with high semantic relevance, ensuring enhanced recommendation accuracy. Currently, PersonalityBERT is optimized for book recommendations, but its applications can be extended to various domains, such as movie recommendations on Netflix, career guidance platforms, and AI-driven personalized learning systems. This research demonstrates that deep learning-based personality prediction models using BERT provide accurate and meaningful recommendations. Future work involves further semantic integration, cross-domain applications, and expanding the model to multilingual personality analysis to improve recommendation diversity and accuracy. Supervised by Dr. Asfand-e-Yar Wed, 01 Jan 2025 00:00:00 GMT http://hdl.handle.net/123456789/19845 2025-01-01T00:00:00Z http://hdl.handle.net/123456789/19846 A Multi-Modal Video Recommendation Approach for Cold-Start Problem Fawad Ur Rehman, 01-249231-005 Rapid growth of short-video content on platforms such as TikTok, YouTube, Facebook, and Instagram has intensified the demand for robust recommendation systems (RS) that can effectively handle the cold-start problem. Traditional methods, such as collaborative filtering (CF) and content-based (CB) approaches, struggle in these scenarios due to their reliance on historical interaction data, which are often unavailable or sparse for new users and items. To address these challenges, this thesis proposes a novel multimodal recommendation framework that integrates textual, visual, aural, and metadata features using advanced deep learning models to enhance recommendation accuracy and mitigate cold-start limitations. The proposed framework employs a hybrid architecture featuring modalityspecific encoders that extract rich representations from different content types. A Graph Neural Network (GNN)-based fusion encoder models relationships between items by aggregating features across modalities, while mutual information maximization ensures alignment between latent representations and raw inputs, improving cross-modal consistency. Additionally, a generative decoder reconstructs the original features to preserve semantic fidelity, enabling robust latent space learning even in sparse interaction scenarios. Empirical evaluation of the MicroLens dataset, a large-scale benchmark for short-video recommendations, demonstrates that the proposed framework outperforms state-of-the-art baselines in both standard and cold-start conditions. These findings highlight the potential of multimodal learning to bridge the gap between sparse interaction data and rich content representations, offering practical insights for social media platforms and streaming services aiming to improve user engagement and retention. Supervised by Dr. Fatima Khallique Wed, 01 Jan 2025 00:00:00 GMT http://hdl.handle.net/123456789/19846 2025-01-01T00:00:00Z http://hdl.handle.net/123456789/19844 Predicting Efficient Outcomes in Trademark and Copyright Cases using NLP Methods Roheen, 01-249231-017 Natural Language Processing (NLP) in the legal domain has been a vibrant area of research for many years, while the ability to process text has effectively increased with the development of AI and NLP techniques. Due to the increasing number of court proceedings, particularly those related to Intellectual Property Rights (IPR) in Pakistani judiciary, it is difficult and time-consuming for lawyers to navigate and extract valuable insights from legal data. Thus, there is a growing need for an efficient legal assistance system that can provide major improvements in the efficiency of court procedures. A novel semantic search engine is designed in this research to assist lawyers in managing and drafting IPR cases and extract relevant legal data. This search assistance system can predict court judgments and extract relevant data from Trademark and copyright cases as well as Ordinance Data based on the user’s input query. For judgment forecasting of legal scenarios, XGBoost, SVM, Random Forest (RF) were used, with the mean cross-validation score as 75%, 88%, and 91%. The use of pre-trained BERT model in the designed system further enhances the efficiency of data retrieval. In terms of cases and ordinance data extraction, the Mean Average Precisions (MAP) of PAK-LEGAL-BERT and legal-bert-base ranges between 67% to 71%. The models are then fine-tuned on domain-specific data and then used to extract relevant data, thus MAP values increase from 85% to 95% effectively. Supervised by Dr. Asfand-e-Yar Wed, 01 Jan 2025 00:00:00 GMT http://hdl.handle.net/123456789/19844 2025-01-01T00:00:00Z http://hdl.handle.net/123456789/20736 Forecasting the Dynamics of Cryptocurrency Rates by Inspecting Technical Patterns Abdul Wahab, 01-249232-001 The unprecedented growth and extreme volatility of the cryptocurrency market have created an urgent and persistent need for reliable and precise forecasting models that can assist investors, traders, and financial analysts in making well-informed and timely decisions. Cryptocurrency prices are influenced by a variety of factors, including market sentiment, macroeconomic trends, trading volume, and technological developments, making prediction a complex and multidimensional task. Traditional statistical models often fall short in capturing the nonlinear dependencies and temporal patterns inherent in such data, highlighting the necessity for advanced approaches. This research introduces a novel hybrid deep learning framework that com- bines the strengths of Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) neural networks to accurately forecast price movements for ten major cryptocurrencies: Bitcoin (BTC), Ethereum (ETH), Dogecoin (DOGE), Shiba Inu (SHIB), Ripple (XRP), NEAR Protocol (NEAR), SUI20947, PEPE24478, and Solana (SOL). By integrating both short-term high-frequency intervals (five-minute price data) and long-term daily closing prices, the proposed methodology is designed to capture a diverse range of market dynamics, seasonal trends, and high-volatility fluctuations. To enhance predictive accuracy and practical usability, the framework incorporates widely used technical indicators, including the Relative Strength Index (RSI), Simple Moving Average (SMA) crossovers, and Bollinger Bands. These features are merged with deep learning outputs to produce predictions that are both statistically sound and aligned with real-world trading strategies. The dataset, sourced from Yahoo Finance, undergoes preprocessing steps such as normalization and trans- formation using a sliding window approach to enable the generation of multi-step forecasts. Model performance is rigorously evaluated using the Mean Absolute Percent- age Error (MAPE) metric. When applied to long-term (daily) price data, the LSTM- GRU hybrid model achieves a forecast accuracy in the range of 95% to 96%. In contrast, when tested on short-term (five-minute) interval data, the GRU-based implementation alone yields an exceptional accuracy of approximately 99%. These results underscore the adaptability and robustness of the proposed approach across multiple timeframes. The findings demonstrate that the integration of deep learning architectures with domain-specific technical indicators results in a more reliable, generalized, and scalable solution for cryptocurrency price forecasting. The cross-asset, cross- timeframe experimental design offers a practical blueprint for real-time algorithmic trading systems and provides a foundation for the development of adaptive financial models. Furthermore, the study emphasizes the importance of leveraging temporal features, technical patterns, and advanced neural architectures to improve decision making capabilities in the inherently volatile and unpredictable digital asset markets. Supervised by Dr. Faryal Nosheen Wed, 01 Jan 2025 00:00:00 GMT http://hdl.handle.net/123456789/20736 2025-01-01T00:00:00Z