Personality-based Recommender System
Lead Research Organisation:
University of Cambridge
Department Name: Computer Science and Technology
Abstract
This project proposes a novel method for building a recommender system, which would combine non-parametric Bayesian models using user personality and item information to generate item suggestions.
Non-parametric Bayesian models (NPBM) are statistical methods in which the amount of captured information increases with the size of the data set. In contrast to parametric models, which use a fixed number of parameters to represent the way the data was generated; NPBM make weaker assumptions as their prediction complexity increases with the number of data samples. Naturally, representing a finite number of training samples with a model, which has an infinite number of parameters, would result in capturing specific properties of training data rather than generalizations of the entire data set, known as 'overfitting'. However, Bayesian approaches are not susceptible to this problem since they average over parameters to estimate their probability rather than fitting them to the data. These properties make the NPBM highly flexible, with better predictive performance and more realistic data representation.
Deep Learning algorithms attempt to attempt to model multiple levels of representations with increasing abstraction. The advantage of deep learning algorithms over traditional machine learning is that they generate distributed data representations, meaning that they form non-mutually exclusive features to perform learning tasks. This property increases their generalization capabilities because it enables them to operate on unseen examples.
A Recurrent Neural Network(RNN) is a deep learning algorithm, which can operate on arbitrary long sequences, hence it will be used to learn personality characteristics from user text. These characteristics will be clustered into personality types, using the Chinese restaurant process (CRP), a type of NPBM which models the probability distribution over partitioning the users into possibly infinite number of personality types. Further, Indian Buffet Process, an extension to the CRP, which may assign a single data point to multiple partitions, will model hidden item features and determine item similarity based on the number of overlapping features, to which items are assigned. Thus, it will be possible to not only investigate the resulting performance of the recommender system, but also identify possible qualities of the items, which are important for different personality types.
Commercial applications of the project are possible within large-scale e-commerce and recommender website such as Amazon, eBay, IMDB, which are constantly investigating more efficient methods of suggesting items of interest to their users. Target groups for advertising campaigns can be elicited more efficiently and effectively due to the fact that they will be formed automatically on the basis of innate human characteristics.
Non-parametric Bayesian models (NPBM) are statistical methods in which the amount of captured information increases with the size of the data set. In contrast to parametric models, which use a fixed number of parameters to represent the way the data was generated; NPBM make weaker assumptions as their prediction complexity increases with the number of data samples. Naturally, representing a finite number of training samples with a model, which has an infinite number of parameters, would result in capturing specific properties of training data rather than generalizations of the entire data set, known as 'overfitting'. However, Bayesian approaches are not susceptible to this problem since they average over parameters to estimate their probability rather than fitting them to the data. These properties make the NPBM highly flexible, with better predictive performance and more realistic data representation.
Deep Learning algorithms attempt to attempt to model multiple levels of representations with increasing abstraction. The advantage of deep learning algorithms over traditional machine learning is that they generate distributed data representations, meaning that they form non-mutually exclusive features to perform learning tasks. This property increases their generalization capabilities because it enables them to operate on unseen examples.
A Recurrent Neural Network(RNN) is a deep learning algorithm, which can operate on arbitrary long sequences, hence it will be used to learn personality characteristics from user text. These characteristics will be clustered into personality types, using the Chinese restaurant process (CRP), a type of NPBM which models the probability distribution over partitioning the users into possibly infinite number of personality types. Further, Indian Buffet Process, an extension to the CRP, which may assign a single data point to multiple partitions, will model hidden item features and determine item similarity based on the number of overlapping features, to which items are assigned. Thus, it will be possible to not only investigate the resulting performance of the recommender system, but also identify possible qualities of the items, which are important for different personality types.
Commercial applications of the project are possible within large-scale e-commerce and recommender website such as Amazon, eBay, IMDB, which are constantly investigating more efficient methods of suggesting items of interest to their users. Target groups for advertising campaigns can be elicited more efficiently and effectively due to the fact that they will be formed automatically on the basis of innate human characteristics.
Organisations
People |
ORCID iD |
Mateja Jamnik (Primary Supervisor) | |
Botty Dimanov (Student) |
Publications
Studentship Projects
Project Reference | Relationship | Related To | Start | End | Student Name |
---|---|---|---|---|---|
EP/N509620/1 | 30/09/2016 | 29/09/2022 | |||
1778323 | Studentship | EP/N509620/1 | 30/09/2016 | 30/03/2020 | Botty Dimanov |
Description | The findings have attracted serious interest from industry partners and investors and will probably lead to a spin-out company. |
Sector | Digital/Communication/Information Technologies (including Software) |
Impact Types | Economic |
Title | Step-wise Sensitivity Analysis |
Description | We introduce a novel method for interpreting Deep Neural Networks (DNN) clas- sification decisions. Our method constructs a dependency graph of the relevant neurons across the network to gain a fine-grained understanding of the nature and interactions of DNN's internal features. |
Type Of Material | Technology assay or reagent |
Year Produced | 2018 |
Provided To Others? | Yes |
Impact | Interest from industry and investors. |
Description | i-teams - A project aimed at helping researchers commercialise their projects. |
Form Of Engagement Activity | A formal working group, expert panel or dialogue |
Part Of Official Scheme? | No |
Geographic Reach | International |
Primary Audience | Industry/Business |
Results and Impact | It is a 10 week course involving customer validation, speaking to industry leaders and understanding their needs. We have identified more than 10 companies, which are interested in the research and further collaboration. |
Year(s) Of Engagement Activity | 2018,2019 |