Brief overview
Duration: from 5 hours (incl. lunch break)
Contents: Data Science, scalar product, vectors, matrices, Machine Learning
Previous knowledge: Function concept, differential calculus, linear systems of equations
Participants: Mathematics courses from grade 10
Created by: Sarah Schönbrodt
Registration: Appointments can be made individually using this form.
Netflix, Amazon, Zalando and others are focusing on one thing in particular when it comes to customer loyalty: good recommendations for new products, films, clothing, etc. that are tailored to the user. To this end, these companies develop recommendation systems that can predict as well as possible what the respective user might like.
To further improve its own recommendation system, Netflix set up a challenge in 2006: The team that could predict at least 10% more accurately which movies a user would like had a chance to win the grand prize of $1 million!
In the workshop you will work with the original data from the Netflix Challenge and develop your own recommendation system. You will apply common strategies from the field of artificial intelligence.
Timetable
Phase | Topic | Content | School reference | further math. Contents | media/ materials | Time (min.) |
Entry + Technique introduction | The problem | Getting started with recommender systems and the Netflix Challenge | - | - | Presentation slides | 15 10 |
Elaboration | Data Analysis | Introduction and overview of the Netflix data set. | Understanding data, mean, standard deviation. | - | AB1-SuS | 30 - 40 |
Backup | Discuss findings from analysis of data set | Presentation slides | 15 | |||
Development | The model | Recognize correlations in valuation data, work out idea of factorization | Understanding of data, vectors, scalar product | matrices, matrix multiplication | AB2-SuS | 30 |
Backup | Discussion of captured relationships in the data, backup of the factorization idea. | Presentation slides | 10 | |||
Elaboration | Error functions | Determination of a decomposition by hand, evaluation of the quality of a found decomposition (factorization) | Sum of the error squares | AB3-SuS | 30 - 40 | |
Backup | Presentation slides | 10 | ||||
Elaboration | The optimization | Development of an optimization procedure to determine a factorization
| derivatives, minimization problem, linear system of equations | Functions in two variables, gradient | AB4-SuS | 50 - 60 |
Elaboration | The optimization | Application of the optimization procedure from AB 4 (long version | AB4-short-SuS | 20 | ||
Backup | - | - | Presentation slides | 10 | ||
Elaboration | Application to Netflix dataset | Understanding of training and test data
| Interpretation and critical reflection of the results | Principle of machine learning | AB5-SuS | 20 - 25 |
Backup + Final discussion | Critical discussions | Comparison of the obtained results on the Netflix dataset. Social discussion of the use of recommender systems. | - | - | Presentation slides | 20 |
Additional material | ||||||
Additional sheet: Scalar product (linked to AB2) | Mathematical reasoning for larger or smaller predicted valuations | Scalar product, cosine, vectors, angles | - | AB_SKP-SuS | ||
Additional sheet: error measures | Comparison of different error functions | Functions / function graphs | Optimization: Properties of the objective function | AB_error-mass-SuS | ||
Additional sheet: Regularization (linked to AB4) | Regularization - avoiding overfitting to the training data | Graphs of functions | Optimization: existence and uniqueness of a solution | AB_Regularization-SuS | ||
Additional sheet: own recommendations (linked to AB 5) | Give own evaluations and calculate individual recommendations | - | - | AB_myRecommendations-SuS | ||