Recommendation in Context

1
Recommendation in Context

• Sarabjot Singh Anand
• University of Warwick
• s.s.anand_at_warwick.ac.uk

2
Defining Context

• Entities interact with their environment through
  situated actions
• Context may be defined as any information that
  can be used to characterise the situation of
  entities (Day et al., 2001)

3
Properties of Context

• Context is
• not observable though behaviour arising from it
  is
• not static
• stochastic process with d states c1,c2,,cd
• What is the value of d?
• That which maximises the likelihood of the
  observed user behaviour
• For a system to be context aware
• Discovering context may not be as important as
• recognising behaviour arising from the context
• adapting to the needs of the user within the
  context

4
User Model

• Short Term Memory (STM)
• Current interaction with the user
• Modelled as a rating vector
• where, I is the set of Itemsa?U is the active
  user
• Long Term Memory (LTM)
• ?ci,rai? ci?C, rai?R
• Where R is the set of all rating functions

5
Recommendation Generation

• Context Identification
• Context in implicit in the observable user
  behaviour
• Discover fU?C
• Rating Prediction
• Use f(ua) to predict ratings for each item not
  rated by ua
• Recommend items with highest predicted ratings

6
Memory Interaction Models

• Inclusive Memory Model
• Uses all ratings in the LTM and STM of ua to
  define neighbourhood
• Temporal Memory Model
• Uses ratings from STM and the last k ratings from
  LTM
• Contextual Memory Model
• Uses ratings from STM and those ratings from LTM
  rated within the same context as current context

7
Example Implementations

• Assumption
• context does not change within a user visit
• model
• Separate rating function for each visit
• A visit consists of a timestamp and a rating
  function
• Algorithms
• STO Uses ratings from the current visit of ua
  only
• SLLT Uses ratings from current visit and the
  users previous visit
• SFLT Uses all ratings
• SCLT Uses ratings from current visit and ratings
  from those visits with similarity to STM gt
  sim_threshold

8
i1 i2 i3 .im
u1 u2 . . . . un
if val.vas gtsim_threshold, we assume visits val
and vas have the same context
9
Results Effect of Similarity Threshold

• STM has a positive effect on Recall
• When the number of items rated in STM is large,
  LTM can add noise to the recommendation process
• Optimal value of Similarity threshold is 0.9
  compared with 0.3 for small STM

10
Results Algorithm Performance
Note ltalgorithmgt0 denotes the use of algorithm
where only long term memory is used
11
Item Knowledge Bases and Context

• Given user behavioural data and an item knowledge
  bases
• Discover different user behaviours that can be
  associated with different user interaction
  contexts

12
A High Level View of our Objective
Ontological Profile Generator
0.3
1
0.5
0.2
0.75
0.05
0.15

• One visitor may have multiple such profiles
• If they are distinct enough, they would represent
  a different context for the user visit
• Clustering of these profiles using CobWeb with
  cross validation to choose the number of clusters
  identified 27 distinct clusters (contexts) within
  15,000 user visits

13
Generalized Cosine Max Similarity Metric
Similar of ij and it depends on the context of
the users visit
14
Calculating S
15
Measuring Item Similarity

• Object Similarity
• Inter-class similarity
• Similarity between the classes of the objects
• Ratio of the number of attributes in common to
  the total number of unique attributes
• Intra-class similarity
• Attribute distance for attributes that the
  objects have in common
• A documentary and a fiction movie have a title,
  director, length etc. in common
• Aggregation function

16
Complexity of Item Similarity Measurement

• Algorithm is in O(md)
• m is the sum of the cardinalities of the
  relationships between the item and other concepts
  used in describing it
• d is the depth to which the item is described

17
Understanding Impact

• If Tom has rated highly, the following movies
• Is he a M. Night Shyamalan fan or a Bruce Willis
  fan?
• Would Signs be a good recommendation or would Die
  Hard be a better recommendation?
• Impact measures the effect that a particular
  concept, describing an item, has on item
  selection

18
Measuring Impact

• Defined based on an observed (f(x)) and an
  expected distribution (g(x)) of instances of the
  concept
• The greater the divergence (Kullback-Leibler)
  between these distributions, the greater the
  impact
• impu assumes g(x) to be uniform
• All instances on the concept are equally likely
  to be viewed by the user
• s is the number of unique instances of the
  concept and H(f(x)) is the entropy of f(x)

19
Measuring Impact (II)

• impI assumes g(x) is the likelihood of instances
  of the concept being viewed within a random
  sample
• Simulated
• using the item knowledge base, assuming each item
  has an equal probability of being selected
• Popularity of the item across all users (takes
  temporal factors into account)

20
Evaluation Results
21
Conclusions

• Evidence suggests that incorporating context
  within recommendation generation can improve
  accuracy of recommender engines
• Availability of item knowledge bases can help
  identify underlying user behaviours
• Modelling and incorporating context within
  recommendation is in its infancy
• Knowing what the context isnt as important as
  knowing that
• different contexts exist
• observable behaviours is conditioned on the
  underlying context