Serendipity module in Item Recommender System

Serendipity module in Michele Filannino
[info@ﬁlanninomichele.com]

ITem Recommender Piero Molino
[piero.molino@gmail.com]

Outline

• What is serendipity

• Computer science applications

• Serendipity in the Information
Filtering

• Evaluation metrics

• A new approach

• Architecture

• Upcoming developments

Origins [1]

• Christopher Armeno, 1557
Persian fairy tale “The Three
Princes of Serendip”

• Horace Walpole, 1754
“Making discoveries by
accident and sagacity, of things
which one is not on quest of”
ispiring from fairy tale

• Pek van Andel, 1994
“The art of making an unsought
ﬁnding”

Discoveries and inventions

• The discovery of America by Christopher Columbus

• Gelignite by Alfred Nobel, when he accidentally mixed collodium (gun cotton)
with nitroglycerin

• Penicillin by Alexander Fleming. He failed to disinfect cultures of bacteria
when leaving for his vacations, only to ﬁnd them contaminated with
Penicillium molds, which killed the bacteria

• The psychedelic effects of LSD by Albert Hofmann

• Cellophane, was developed in 1908 by swiss chemist Jacques
Brandenberger, as a material for covering stain-proof tablecloth

• The structure of benzene by Friedric August Kekulé.

Serendipity in
scientiﬁc research [2]

• "It should be recognized that
serendipitous discoveries are of
signiﬁcant value in the advancement of
science and often present the foundation
for important intellectual leaps of
understanding" - M.K. Stoskopf

• Serendipity as result of a wide culture and
curious open mind

• These characteristics allow to recognize
serendipity on manifestation

Serendipity, creativity and randomness [1] [3]

• Seredipity don’t come from randomness, but from events brought to light by
an activity on the edge between consciousness and unconsciousness

• Classiﬁcation of “information seekers”

• Role of personal characteristics in the serendipity

• “Lateral thinking” and methods (de Bono)

Serendipity equations 1/2 [4]

• P = Problem
KP = Knowledge domain
EP = Incorrect knowledge
M = Inspiring metaphor
KM = Knowledge domain of inspiring metaphor
S = Solution
KN = Additional knowledge

• Conventional creativity
P1 ∈ (KP1), M ∈ (KM) S1 ∈ (KP1, KM, KN)

• Serendipity
P1 ∈ (KP1), M ∈ (KM) P2 ∈ (KP2), S2 ∈ (KP2, KM, KN)

Serendipity equations 2/2 [4]

• P = Problem
KP = Knowledge domain
EP = Incorrect knowledge
M = Inspiring metaphor
KM = Knowledge domain of inspiring metaphor
S = Solution
KN = Additional knowledge

• Serendipity without inspiring metaphor
P1 ∈ (KP1) P2 ∈ (KP2), S2 ∈ (KP2, KN)

• Serendipity from incorrect knowledge
P1 ∈ (KP1, EP1) P2 ∈ (KP2), S2 ∈ (KP2, KN)

Computer Science
applications

Max 1/2 [5]

• Software agent that browse the web, simulating human behaviour, searching
for interesting pieces of information

• The target is to encourage user creativity allowing new access points to
informations and to lead serendipity-based discoveries

• It uses IR methods and ad-hoc heuristics

Max 2/2 [5]

• Search and browse process (best ﬁt - treshold)

• External product based heuristic valutation

• Interaction with users happens via e-mail

• It uses WordNet

Serendipity in the
Information Filtering

Yesterday, today and tomorrow...

• “Informations research with computer technologies may tend to reduce the
opportunity for serendipitous informations encounter” - Gup (1997-1998)

• “The image of the academic specialist, searching the shelves for a
serendipitous connection, may seem quaint, but it remains powerful. The
challenge for the digital library is to preserve this opportunity in cyberspace” -
Huwe 1999

• There is a level of emotional reaction associated with serendipity that is
difﬁcult to capture in any metric

Obviousness [6]

• Examples (Travels - White
Album - Star Trek)

• Ratability: probability that an
item will be the next item that
the user will consume (and then
rate) given what the system
knows of the user’s proﬁle

• The implicit assumption is that
a user is always interested in
the items with the highest
ratability. This assumption is
true in classiﬁcation problems,
isn’t true in recommenders

Novelty vs Serendipity [7] [8]

• Both examples of not-obviousness

• Novelty: A serendipitous recommendation helps the user ﬁnd a surprisingly
interesting item he might have autonomously discovered

• Serendipity: A serendipitous recommendation helps the user ﬁnd a
surprisingly interesting item he might not have otherwise discovered

• Example: Movie Recommender

Inadequacy of classic
metrics [7]

• Classic metrics don’t take into
account obviousness, novelty
and serendipity

• Accurate reccomendation ≠
Useful reccomendation for user

• It’s impossible to evaluate the
serendipity degree without
considering user feedback

User-based evaluation [9]

• Users don’t want an algorithm with best score, but a sensible recomendation

• We need to consider the tasks and targets of the user in relation with different
algorithms to obtain useful reccomendation

• Human-Recommender Interaction: Framework to structure interaction
aspects between human and recommender, based on user experience and
needs

Suggestions

• Interview user with:

• Unknown items percentage
in respect to all the articles
suggested

• Interesting items
percentage in respect to all
the articles suggested

• Satisfaction about the
reccomandation

Strategies to improve serendipity [3]

• “Blind Luck”: return of casual reccommendation

• “Prepared Mind”: deep user proﬁling

• “Anomalies and Exception”: search by poor similarity

• “Reasoning by Analogy”: not implemented yet

Base assumption

• The user profile doesn’t represent user tastes like in a classic recommending
system, but it represents what the user knows

• The user profile can be updated with informations not only about the
purchased items, but also about researches because, if the user searches for
something, this is known or it become known after the showing of the search
results

• The user profile can be updated also with informations about the item
visualization

Probability of serendipitous happenings

• Serendipity can’t happen if the user already knows what is recommended

• The lower is the probability that the user knows an item, the higher is the
probability of a serendipitous reccomendation.

• We can say that te probability that the user knows something semantically
near to what we are sure he knows is higher than the probability of something
semantically far

• If we evaluate semantical distance with a similarity metric, it results that is
more probable to get a serendipitous reccomandation recommending to the
user something less similar to his proﬁle

To support, not to substitute

• A proposal with the intent to promote serendipity can be based on poor
similarity

• Obviously in the practical use of a recommender we can’t entrust only to
serendipitous reccomandations

• It’s possible to support a reccomandation based on classic methods with a
serendipitous reccomendation that stimulates the user and gives him new
entry points to the items in the system

Noble and practical objectives

• The objectives are:

• Noble: to enable the user to know something new, to make intresting
discoveries, to find something different from what he is used to, stimulatig
his curiosity

• Practical: to improve the possibility that the user knows an item that he
couldn’t otherwise have known (or that he would have been difficult to find
otherwise), to improve the overall serendipity rate of the system
reccomendations

Knowledge profile

• The user profile usually represents the user’s tastes

• A profile that represents the user’s knowlege, the areas of interest, and so on
would be more useful for the implementation of a serendipity module

• For that reason it would be usefull to track the page visits and the serches
made by the user

Inverted profile

• The aim is to search by poor similarity, so the system will create an “inverted”
version of the user profile

• Let’s substitute the tf-idf weights with new weights obtained by this formula:

• ∀ wi ∈ P: nwi = maxweight(P) - wi

• wi is the weight of the word in the i-th position in the original vector,
maxweight(P) is the highest weight in the profile P, nwi is the weight of the
word in the i-th position in “inverted” vector

Randomness inside the treshold

• To avoid cold start problems and repeated recommendations it’s possible to
select a random reccomendation

• Given a list of results ordered by ranking, it’s possible to set a treshold of
similarity (poor similarity) and to select a random item to reccomend inside
this treshold

• It’s also possible to select a random item to reccommend from the x more
similar (poorly similar) regardless of the similarity (poor similarity). The x is
propotional to the total number of items in the system

Structure

ITR

Serendipitous
Profile
Reccommendation

Serendipity Module

Query Inverse Profile
Results Generator

Inverted
Profile
Query

Recommendation
Generator

Future developments

• Implementation of a Reasoner by Analogy

• Implementation of the other algorithms proposed by de Bono and Toms

• Implementation of a system that selects which algorithm to use depending on
the kind of the user and his task

• Design of a “virtual shopkeeper” to interact with while browsing that analizes
the user and his task and place them inside a HRI proﬁle and changes the
systeam consequently, changing the retrieval algorithm, the ﬁltering approach,
etc.

Bibliography 1/2

• [1] Anatomy of the Unsought Finding. Serendipity: Orgin, History, Domains,
Traditions, Appearances, Patterns and Programmability - van Andel (1994)

• [2] Serendipity and Information Seeking - Foster & Ford (2003)

• [3] Serendipitous Information Retrieval - Toms (2000)

• [4] The Serendipity Equations - de Figueiredo, Campos (2001)

• [5] Searching the Unsearchable: Inducing Serendipitous Insights - de
Figueiredo, Campos (2001)

Bibliography 2/2

• [6] Accurate is not always good: How Accuracy Metrics have hurt
Recommender Systems - McNee, Riedl, Konstan (2006)

• [7] Evaluating Collaborative Filtering Recommender Systems - Herlocker,
Konstan, Terveel, Riedl (2004)

• [8] Modern Information Retrieval - Baeza-Yates, Ribeiro-Neto (1999)

• [9] Making Recommendations Better: An Analytic Model for Human-
Recommender Interaction - McNee, Riedl, Konstan (2006)

Serendipity module in Item Recommender System

More Related Content

Similar to Serendipity module in Item Recommender System (20)

More from Michele Filannino (15)

Recently uploaded (20)

Serendipity module in Item Recommender System