Pavlos Andreadis and Paul Henderson
Thu 02 Mar 2017, 12:45 - 13:45

If you have a question about this talk, please contact: Magda Nowak (mnowak)

pastries will be served

Pavlos Andreadis  "Efficient Preference Elicitation with Coarse Preferences"

"We present a new model for the representation of preferences that are based on coarse criteria, leading to better recommendations in problems where user behaviour is consistent with them making decisions  by  aggregating  many  individual  options. When  this  is  the  case,  our  'coarse  preference' model enables an elicitation procedure that outperforms  the  state-of-the-art  utility  elicitation  procedure  that  is  based  on  the  assumption  of  additive independence  of  variables.   We  demonstrate  this through experiments performed with a real-world data  set  derived  from  a  mobile  clothes  shopping app.   We show that our proposed model achieves double the recommendation acceptance rate and a significant  increase  in  platform  profit,  due  to  the use of a decomposition that is better suited to the true nature of the users' indifference regions.  Furthermore,  our  model  is  of  lower  complexity  than similar  additive  independence  models,  with  computation  time  rising  at  a  slower  rate  with  respect to dimensionality of item descriptions - achieving a halving of computation time in our experiment. Our results have insights for recommendation systems designers,  potentially enabling new ways of representing  user  utility  when  secondary  criteria need to be maximised without compromising performance."


Paul Henderson  ''End-to-end training of object class detectors for mean average precision''

I will present a method for training CNN-based object class detectors directly using mean average precision (mAP) as the training loss, in a truly end-to-end fashion that includes non-maximum suppresion (NMS) at training time. mAP following NMS forms a piecewise-constant structured loss over thousands of windows, with gradients that do not convey useful information for gradient descent. Hence, we define new, general gradient-like quantities for piecewise constant functions, which have wide applicability. We describe how to calculate these efficiently for mAP following NMS, enabling us to train a detector based on Fast R-CNN directly for mAP.