Information market based recommender systems fusion Efthimios - - PowerPoint PPT Presentation

information market based recommender systems fusion
SMART_READER_LITE
LIVE PREVIEW

Information market based recommender systems fusion Efthimios - - PowerPoint PPT Presentation

Information Management Unit / ICCS of NTUA imu.iccs.gr Information market based recommender systems fusion Efthimios Bothos Konstantinos Christidis Dimitris Apostolou National Technical National Technical University of Piraeus University


slide-1
SLIDE 1

Information Management Unit / ICCS of NTUA imu.iccs.gr

Information market based recommender systems fusion

Gregoris Mentzas National Technical University of Athens Greece Efthimios Bothos National Technical University of Athens Greece Konstantinos Christidis National Technical University of Athens Greece Dimitris Apostolou University of Piraeus Greece

Information Management Unit, HetRec Workshop, RecSys October 23-27, 2011 Chicago, IL USA 1

slide-2
SLIDE 2

Information Management Unit / ICCS of NTUA imu.iccs.gr

Information Management Unit, HetRec Workshop, RecSys October 23-27, 2011 Chicago, IL USA 2

Outline

 Introduction  Background on Information Markets  Information market based recommender systems fusion  Experiments and Results  Conclusions and Further Work

slide-3
SLIDE 3

Information Management Unit / ICCS of NTUA imu.iccs.gr

Information Management Unit, HetRec Workshop, RecSys October 23-27, 2011 Chicago, IL USA 3

Research Area: Ensemble Recommenders

 Recent work in recommender systems proposes the

acquisition of results by combining sets of recommendation models

  • Aggregation of the predictions of different base algorithms - the

ensemble - to obtain a final prediction

 The combination of different predictions into a final prediction is also

referred to as blending or fusion

  • E.g. the models proposed for the NetFlix Prize have been combined to

address the problem of recommendation in the specific dataset

 Most existing methods presuppose restrictive assumptions

  • Most ensembles of recommender s have a constant composition
  • The training data performance has to be a good proxy for subsequent

actual performance

 Cannot easily adapt to changes in future user behavior

slide-4
SLIDE 4

Information Management Unit / ICCS of NTUA imu.iccs.gr

Our Approach

 In this paper we use the market paradigm for blending

recommender systems

 Market participants are computational agents

  • Representing different base recommenders

 Agents invest/bet on the recommendation they foresee to

be correct

  • Based on the information provided by their corresponding base

recommender

 The recommendation is based on the market outcome

  • It depends on the wealth of the participants and reflects the „wealth-

weighted opinions' of the base recommenders

Information Management Unit, HetRec Workshop, RecSys October 23-27, 2011 Chicago, IL USA 4

slide-5
SLIDE 5

Information Management Unit / ICCS of NTUA imu.iccs.gr

Information Management Unit, HetRec Workshop, RecSys October 23-27, 2011 Chicago, IL USA 5

Outline

 Introduction  Background on Information markets  Information market based recommender systems fusion  Experiments and Results  Conclusions and Further Work

slide-6
SLIDE 6

Information Management Unit / ICCS of NTUA imu.iccs.gr

Information Management Unit, HetRec Workshop, RecSys October 23-27, 2011 Chicago, IL USA 6

Information Markets (IMs)

In general markets provide mechanisms for risk sharing or resource allocation

  • Market prices are able to aggregate and convey information
  • Efficient markets hypothesis: market in which prices always “fully reflect”

available information is called “efficient” (Fama, 1970)

IMs are markets designed and run for the primary purpose of mining and aggregating information scattered among participants

  • Also known as Prediction markets, Decision markets

IMs make use of specifically designed contracts that yield payments based on the outcome of uncertain future events

  • Contrary to traditional equity markets contracts are not tied to a claim of an
  • wnership stake in a firm
  • The assets are claims that will pay off an amount which depends upon the

state of the world

slide-7
SLIDE 7

Information Management Unit / ICCS of NTUA imu.iccs.gr

An Example

Information Management Unit, HetRec Workshop, RecSys October 23-27, 2011 Chicago, IL USA 7

Will United States GDP growth for Q3 of 2011 be positive?

  • Traders who believe it will be positive buy contracts, otherwise

sell contracts

Will be announced today, but since price reflects probability we are pretty confident the GDP will be positive!

slide-8
SLIDE 8

Information Management Unit / ICCS of NTUA imu.iccs.gr

Information Management Unit, HetRec Workshop, RecSys October 23-27, 2011 Chicago, IL USA 8

IMs Applications

IMs with human participants have done well in various contexts:

  • Predictions

 Iowa Electronic Markets beat US presidential election polls 451/596 (Berg et al.

2008)

 Oscar winners: Hollywood Stock Exchange beats individual and average forecasts

  • f 5 experts (Lamare, 2007)

 NFL: Markets rank 11th and 12th against 1947 humans (human average 39th)

(Servan-Schreiber et al. 2005)

  • Decision support (opinion polling)

 Hewlett-Packard market beats official forecasts in 6 out of 8 events (Chen & Plott

2002)

  • Preferences

 Selection of product concepts with IAMs provides similar results with surveys

(Dahan et al. 2010)

IMs with computational agents have also done well:

  • eg. Predicting the Oscar awards (Bothos et al. 2010)
slide-9
SLIDE 9

Information Management Unit / ICCS of NTUA imu.iccs.gr

Information Management Unit, HetRec Workshop, RecSys October 23-27, 2011 Chicago, IL USA 9

Design Elements

 Contracts reveal participants expectations regarding the

status of future events

  • Payoff depends on the outcome of future events
  • E.g. a futures contract pays $1 if the event occurs, nothing otherwise

 Exchange medium

  • Can be either real or play money

 Trading Mechanisms

  • Define the rules of the market, which specify how orders are placed

and how the price changes

slide-10
SLIDE 10

Information Management Unit / ICCS of NTUA imu.iccs.gr

Design Elements

 Participants bring liquidity and generate the efficient price

  • This means that even uninformed traders must participate. When only

rational traders participate, the “No Trading Theorem” (Milgrom and Stokey, 1982) effect appears and the market cannot function.

  • Participants should be (Surowiecki, 2004)

 Diverse so that people offer different pieces of information  Independent, so that participants pay attention mostly to their own

information, and do not worry about what others think

 De-centralized, so that no one at the top is dictating the crowd's answer

 Openness with respect to information

Information Management Unit, HetRec Workshop, RecSys October 23-27, 2011 Chicago, IL USA 10

slide-11
SLIDE 11

Information Management Unit / ICCS of NTUA imu.iccs.gr

Information Management Unit, HetRec Workshop, RecSys October 23-27, 2011 Chicago, IL USA 11

Outline

 Introduction  Background on Information markets  Information market based recommender systems fusion  Experiments and Results  Conclusions and Further Work

slide-12
SLIDE 12

Information Management Unit / ICCS of NTUA imu.iccs.gr

Our methodology

We consider an IM as an ensemble recommender

  • which can potentially be employed in any recommendation

problem

Information Management Unit, HetRec Workshop, RecSys October 23-27, 2011 Chicago, IL USA 12

  • For a given dataset a set of base

recommenders is considered

  • Base recommenders are trained
  • An IM is composed where its participants are

computational agents representing different base recommenders

  • Agents invest on the rating option they

foresee to be correct

  • They make use of information provided by

their corresponding base recommender

  • When the actual rating is revealed, the agents

that predicted correctly are rewarded

Agent1

Rec1

Agent2

Rec2 Agent3 Rec3

AgentN

RecN

Trained Recommenders Acting as Agents Unrated Item

Information Market

Predicted rating

Recommendation

slide-13
SLIDE 13

Information Management Unit / ICCS of NTUA imu.iccs.gr

Agents

Our agents follow the Belief-Desire-Intention design paradigm

  • According to the BDI framework an agent is characterized by its beliefs, goals

(desires), and intentions

  • The agent intends to achieve his goals given his beliefs about the world

Belief: Stems from the base recommenders

Desire: To maximize their wealth

  • Wealth is determined by the agent‟s forecasting accuracy

Intention: Betting function which defines the wealth percentage an agent will allocate for each rating option

  • Depends on the item to be rated
  • We use a constant betting function: Agents invest independent of the market

price

Information Management Unit, HetRec Workshop, RecSys October 23-27, 2011 Chicago, IL USA 13

N: number of possible options m the agent / recommender m estimatem the otput of the base recommender m Weatlthm :the wealth of agent m

slide-14
SLIDE 14

Information Management Unit / ICCS of NTUA imu.iccs.gr

Information Market

Zero-sum game i.e. the total amount of money collectively owned by the participants is conserved after each new item is presented

Prices denote probabilities of an outcome being correct and sum up to one

Equilibrium price: A unique price that satisfies

  • Total wealth conservation and
  • For constant betting functions solution is provided as follows (Barbu‟10, ICML)

Rewards: The agent is rewarded based on the investment he made on the correct

  • utcome

Information Management Unit, HetRec Workshop, RecSys October 23-27, 2011 Chicago, IL USA 14

k denotes the rating option, x the output of the base recommender, m the agent

slide-15
SLIDE 15

Information Management Unit / ICCS of NTUA imu.iccs.gr

Recommendation and Feedback Process

The expected rating of a new item i for user u is calculated using the current wealth of the agents

Feedback Process

Information Management Unit, HetRec Workshop, RecSys October 23-27, 2011 Chicago, IL USA 15

new item

Market prices are formed Real rating becomes known Wealth is updated Recomme ndation based on current wealth Agents bet

  • n ratings
slide-16
SLIDE 16

Information Management Unit / ICCS of NTUA imu.iccs.gr

Related Work

Wei‟05, Bohte‟04 describe a market-based approach to allocate „consumer attention spaces‟ to recommended items

Resnick‟07, Kutty‟10 use the trading metaphor to limit the influence of manipulators in recommendation problems

Perols‟09 employ information markets to fuse base classifiers in order to identify frauds using financial data

Barbu‟10, Storkey‟11 provide the foundations for using information markets in machine learning and classification problems

In our work we use IMs to build a recommendation system that predicts user opinions on items

Information Management Unit, HetRec Workshop, RecSys October 23-27, 2011 Chicago, IL USA 16

slide-17
SLIDE 17

Information Management Unit / ICCS of NTUA imu.iccs.gr

Information Management Unit, HetRec Workshop, RecSys October 23-27, 2011 Chicago, IL USA 17

Outline

 Introduction  Background on Information markets  Information market based recommender systems fusion  Experiments and Results  Conclusions and Further Work

slide-18
SLIDE 18

Information Management Unit / ICCS of NTUA imu.iccs.gr

Experiment #1 using the Movielens Dataset

The HetRec version of Movielens

  • Contains personal ratings, tags and tag assignments to

movies

  • Ratings range from 1 to 5, including 0.5 steps

Information Management Unit, HetRec Workshop, RecSys October 23-27, 2011 Chicago, IL USA 18

We implemented three base recommenders

  • Collaborative Filtering
  • Content based recommendation using Latent Dirichlet Allocation in order to extract

latent semantics of the tags

  • Average rating of all the users who have already rated the item
  • As a baseline for blending recommenders we used Linear Least Squares

Results based on RMSE

  • The IMs based approach provides

similar results to those of linear least squares regression and

  • Better than those provided by

individual recommenders

slide-19
SLIDE 19

Information Management Unit / ICCS of NTUA imu.iccs.gr

Experiment #2 using the Netflix Dataset

We made use of the publicly available results of Jahrer‟10

  • They randomly split the Netflix probe set with 1.4 million ratings into two

disjoint sets (~700K ratings each)

  • The models were trained on the one set whereas the other is used as a hold
  • ut set

Information Management Unit, HetRec Workshop, RecSys October 23-27, 2011 Chicago, IL USA 19

In this case too

  • The IMs based approach provides similar

results to those of linear least squares regression and

  • Better than those provided by individual

recommenders

slide-20
SLIDE 20

Information Management Unit / ICCS of NTUA imu.iccs.gr

Remarks

 No, our approach didn‟t outperform the linear least squares

ensemble But:

  • Our approach performed close and outperformed all base

recommenders

 Can it adapt to changing conditions?

  • We needed a third experiment…

Information Management Unit, HetRec Workshop, RecSys October 23-27, 2011 Chicago, IL USA 20

slide-21
SLIDE 21

Information Management Unit / ICCS of NTUA imu.iccs.gr

Experiment #3 using a Synthetic Dataset

Using the Movielens dataset, items were divided into three sets

  • Items for which the CF recommender provided the best RMSE then CB and last

averaging rating

  • And arranged in line

We derived least squares factors for the first 20% of the dataset

  • For the latter 80% of the items we applied our market based ensemble and least

squares approach

Information Management Unit, HetRec Workshop, RecSys October 23-27, 2011 Chicago, IL USA 21

The RMSE is lower for the market based ensemble

The wealth of the agents is distributed according to their performance

  • Initially the CF recommender accumulates most
  • f the wealth - the CB approach becomes richer

towards the end of the dataset

CB Avg CF CF

Train Test

slide-22
SLIDE 22

Information Management Unit / ICCS of NTUA imu.iccs.gr

Information Management Unit, HetRec Workshop, RecSys October 23-27, 2011 Chicago, IL USA 22

Outline

 Introduction  Background on Information markets  Information market based recommender systems fusion  Experiments and Results  Conclusions and Further Work

slide-23
SLIDE 23

Information Management Unit / ICCS of NTUA imu.iccs.gr

Conclusions

 We applied an information market-based approach in order to

generate a fusion of recommenders

 Our approach

  • Has comparable performance to linear regression blending methods,

 Experimental results in two datasets have proven that our approach

consistently outperforms the base recommenders and performs similarly with the least squares linear aggregation

  • Does not require offline training data, and
  • Through online learning can adapt to changes in base-recommender

performance

 The proposed approach provides an adaptive framework that can evolve

to address possible changes in recommender predictive power

Information Management Unit, HetRec Workshop, RecSys October 23-27, 2011 Chicago, IL USA 23

slide-24
SLIDE 24

Information Management Unit / ICCS of NTUA imu.iccs.gr

Further Work

 Allow for on-line changes in the ensemble compositions

  • Recommenders can enter or exit the ensemble

 Extend the validation

  • Evaluate the performance of our approach in more datasets

 Especially dataset that contain significant fluctuation in the performance of

the recommendation algorithms

  • Compare with other blending methods

 Examine the performance of other betting functions which

consider the market prices as well

  • This kind of functions will allow our trading agents to learn from each
  • ther as they will monitor price fluctuations, similarly to real life

markets where prices convey information to traders

Information Management Unit, HetRec Workshop, RecSys October 23-27, 2011 Chicago, IL USA 24

slide-25
SLIDE 25

Information Management Unit / ICCS of NTUA imu.iccs.gr

Thank you for listening!