PlayRank Recommendation System

Transcript

1. PLAY RANK By : Erik Gonzalez, Braden Carter, Diego Wearden, Aditya Dhanraj, Krisha Patel, Cristian Cantu
2. How Machine Learning Started
3. 3 •Alan Turing (1950): Proposed the concept of a machine that could simulate any human intelligence task ("Turing Test"). •Hebbian Learning (1949): Donald Hebb introduced a theory on how neurons learn – the foundation of neural networks ("cells that fire together wire together"). •Perceptron (1957): Frank Rosenblatt developed the first neural network model capable of basic pattern recognition. 1940’s – 1950’s
4. Funding and research in AI and ML slowed. Limited computational power and lack of data made it hard to scale ML systems. Early models like decision trees and naive Bayes were introduced but not widely applied. 1970’s -1980’s
5. 1980’s -1990’s •Backpropagation (1986): A breakthrough in training multi-layer neural networks (by Rumelhart, Hinton, and Williams). •Decision Trees & Support Vector Machines (SVMs): Became popular for classification tasks. •Reinforcement Learning (Late 1980s): Algorithms like Q-learning enabled agents to learn through trial and error.
6. 2000’s •The internet explosion provided huge datasets. •Moore’s Law allowed for more powerful processors and GPUs. •Ensemble methods like Random Forests and Gradient Boosting emerged. •Companies like Google and Amazon used ML for recommendations, search, and ads. •Support Vector Machines and kernel methods were dominant in academia.
7. 2010’s •ImageNet (2012): AlexNet, a deep convolutional neural network (CNN), outperformed all competitors and revolutionized computer vision. •Speech & NLP: Deep learning models like RNNs and LSTMs began outperforming traditional methods in speech recognition and translation. •Reinforcement Learning milestones: •AlphaGo defeated world champion in Go (2016). •Generative Models: •GANs (Generative Adversarial Networks) introduced in 2014 (Goodfellow). •Autoencoders became popular for unsupervised learning.
8. 2010’s – 2020’s •Transformer architecture (2017): Introduced by Google in "Attention is All You Need", revolutionized NLP. •BERT (2018) and GPT (2018–2023): Led to language models that could understand and generate text with human-like fluency. •Multimodal models: Models like CLIP and DALL·E can process both images and text. •Scaling Laws: Models trained with more parameters and data performed better (e.g., GPT-3 with 175B parameters, GPT-4 even larger).
9. Recommendation Systems Today
10. 10 1970’s – 1980’s •Early Information Filtering: Basic keyword-based document retrieval systems used in academic libraries and news clipping services. •Expert systems: Rule-based recommendations (e.g., “if a user is interested in X, recommend Y”).
11. 1990’s •1992–1994: Tapestry (Xerox PARC) – One of the first collaborative filtering systems using manual user annotations. •1994: GroupLens – Introduced automated collaborative filtering using user ratings for Usenet news articles. •1997: MovieLens project launched – Still a widely used benchmark dataset today. •1998: Amazon popularizes item-based collaborative filtering for product recommendations.
12. 2000’s •Netflix Prize (2006): $1M competition to improve Netflix’s recommendation algorithm by 10%. Sparked massive innovation in matrix factorization (e.g., SVD, ALS). •Google News Recommendations: Real-time recommender based on collaborative filtering and clustering. •Hybrid Systems: Systems began combining collaborative filtering + content-based filtering.
13. 2010’s •2013: Deep learning enters recommender systems (e.g., Restricted Boltzmann Machines for collaborative filtering). •2015: YouTube publishes deep learning architecture for large-scale recommendations using user watch history. •2016+: Wide & Deep models (Google), DeepFM, Autoencoders, and Neural Collaborative Filtering (NCF). •2017: Reinforcement learning and contextual bandits used for dynamic recommendations (news, ads). •Spotify, TikTok, Netflix: Start using personalized deep neural nets, attention models, and embeddings
14. 2020’s •Transformer-based models (e.g., BERT4Rec, SASRec) introduced for sequential and personalized recommendation. •Graph Neural Networks (GNNs) used to model user–item relationships (e.g., PinSage at Pinterest). •Multimodal recommender systems: Combining text, audio, video, and interaction data. •Reinforcement learning agents: E.g., Netflix uses RL for content selection and adaptive UIs. •Large Language Models (LLMs) begin to support zero-shot and few-shot recommendation (e.g., using GPT to infer preferences). •Privacy-preserving techniques such as federated learning and differential privacy. •Agentic Recommenders: AutoGPT-style systems plan and reason to recommend or automate decision-making.
15. MACHINE LEARNING TOPICS 15 Linear Regression Decision Tree Regression K-Nearest Neighbors Regression Random Forest Regression Gradient Boosting Regression Support Vector Regression
16. 16 WHAT THEY MEAN Linear Regression
17. 17 Decision Tree Regression
18. 18 K-Nearest Neighbors Regression (KNN)
19. 19 Random Forest Regression
20. 20 Gradient Boosting Regression
21. 21 Support Vector Regression (SVR)
22. Regression Models Pros & Cons 22
23. 23 Model Pros Cons Linear Regression Very fast Too simple, cannot handle complex data Decision Tree & Random Forest Can handle yes or no features Puts games into coarse “buckets” with no internal ranking Gradient Boosting Regression Very good at predicting Very slow, needs previous data on what user likes Support Vector Regression (SVR) Can model simple non‑linear relationships via kernels Can only return a score per game KNN Good for classification problems, no need for training. Slow at predicting, need to compare the distances between every other game Pro’s and Con’s
24. KNN Explanation 24
25. 25 How it Looks in Code
26. 26 HYPERPARAMETER TUNING & STACKED ENSEMBLE
30. CONCLUSIO N 30 We used K nearest neighbors' technique as we used the content-based model. The history tab gives the information about the game to the backend to train the model. The features mentioned earlier are what we used! The model calculates based on a similarity score or Euclidean distance ( the distance between two points). And only recommends the top K items returned from KNN.
31. Demo Time !!! 31