Recommendations assumptions
0 likes41 views
This document discusses various types of recommendation systems and assumptions made in developing recommendations. It describes personalized and non-personalized recommendations, which can be user-interaction driven through passive or active feedback, editorially curated, or based on statistics. It also discusses factors that can influence recommendations like similarity, correlations, user activity, sensors, and external influences like mood, time, empathy and induced preferences. It raises questions about using relationships between content, expressing different interest levels, and optimizing the amount of content information presented based on interest, time and mood.
Recommendations assumptions
- 1. + Recommendation Assumptions Prototyping Day @ Mozilla London 3 September 2015 Engineering Summit @ BBC 7 March 2018 Presented at
- 2. + Recommendations types Personal vs. Non personal
- 3. + Recommendations types Personalized User-interaction driven (passive/active) Non personalized Editorially curated Stats Based
- 4. + User-interaction driven passive interaction active interaction A set of user interactions with the system entails a set of (possibly empty) preferences - inference Can be used to reinforce recommendations (like, thumbup, follow, dislike, etc.)
- 5. + Recommendations types Personalized User-interaction driven (passive/active) Non personalized Editorially curated Stats Based
- 8. + Most popular If more than N% + K of people liked, listened to, blablablabed about something (+/- standard deviation), this something is likely to be recommended.
- 9. + Content similarity It can be content-to-content or cross content. To calculate the similarity we can use the actual content or the metadata or both.
- 10. + Co-occurrence relationship (market basket analysis) Understand what products or services are commonly purchased together. If you consume a certain group (cluster) of contents, you are more (or less) likely to consume another group of items Beers and nappies
- 11. + Assumptions Similarity and Correlation User activity assumptions Device sensor analytics (accelerometer, gyroscope, compass, barometer, ambient light sensor, proximity sensor, thermometer, camera, microphone, GPS, etc.)
- 12. + Perturbations (external influences) Model building phase Recommendation phase
- 13. + Mood
- 14. + Wear & Tear
- 15. + Does the time matter?
- 16. + Does the time matter? Are preferences or statistics about a topic valid forever? If not, whats the best time window to take in account? How can we model a preference modification during the time? (Aging, change of taste) How a good/bad feedback can affect a recommendation in the future and for how long?
- 18. + Ive got questions for you Inheritance, taxonomy, membership, how do we use these relationships between contents in a domain in order to infer a user interest? Is the interest of a user black and white only? How can I express different type of interests (if any)? Can we improve the quality of the user experience by presenting the right amount of information of a recommended content according to type and level of interest? time available to consume the content? mood?
- 19. + Thank you Simone.Spaccarotella@bbc.co.uk Senior Software Engineer | BBC Sounds