Tom Rampley Recommendation Engines: an Introduction Slide 2 A
Brief History of Recommendation Engines Today: Recommenders become
core products In addition to Amazon, companies like Pandora,
Stitchfix, and Google (because what is a search engine other than a
document recommender?) make recommendations a core value add of
their services 2000: Amazon joins the party The introduction and
vast success of the Amazon recommendation engine in the early 2000s
led to wide acceptance of the technology as a way of increasing
sales 1992: Recommenders are older than you might think GroupLens
becomes the first widely used recommendation engine Slide 3 What
Does a Recommender Do? Recommendation engines use algorithms of
varying complexity to suggest items based upon historical
information Item ratings or content Past user behavior/purchase
history Recommenders typically use some form of collaborative
filtering Slide 4 Collaborative Filtering The name: Collaborative
because the algorithm takes the choices of many users into account
to make a recommendation Rely on user taste similarity Filtering
because you use the preferences of other users to filter out the
items most likely to be of interest to the current user
Collaborative Filtering algorithms include: K nearest neighbors
Cosine similarity Pearson correlation Bayesian belief nets Markov
decision processes Latent semantic indexing methods Association
Rules Learning Slide 5 Cosine Similarity Example Lets walk through
an example of a simple collaborative filtering algorithm, namely
cosine similarity Cosine similarity can be used to find similar
items, or similar individuals. In this case, well be trying to
identify individuals with similar taste Imagine individual ratings
on a set of items to be a [user,item] matrix. You can then treat
the ratings of each individual as an N- dimensional vector of
ratings on items: {r 1, r 2 r N } The similarity of vectors
(individuals ratings) can be computed by computing the cosine of
the angle between them: The closer the cosine is to 1, the more
alike the two individuals ratings are Slide 6 Cosine Similarity
Example Continued Lets say we have the following matrix of users
and ratings of TV shows: And we encounter a new user, James, who
has only seen and rated 5 of these 7 shows: Of the two remaining
shows, which one should we recommend to James? True BloodCSIJAGStar
TrekCastleThe Wire Twin Peaks Bob5214325 Mary4421312 Jim1152523
George3435543 Jennifer5242410 Natalie0504414 Robin5500422 True
BloodCSIJAGStar TrekCastle James55310 Slide 7 Cosine Similarity
Example Continued To find out, well see who James is most similar
to among the folks who have rated all the shows by calculating the
cosine similarity between the vectors of the 5 shows that each
individual have in common: It seems that Mary is the closest to
James in terms of show ratings among the group. Of the two
remaining shows, The Wire and Twin Peaks, Mary slightly preferred
Twin Peaks so that is what we recommend to James Cosine
SimilarityJames Bob0.73 Mary0.89 Jim0.47 George0.69 Jennifer0.78
Natalie0.50 Robin0.79 Slide 8 Collaborative Filtering Continued
This simple cosine similarity example could be extended to
extremely large datasets with hundreds or thousands of dimensions
You can also compute item to item similarity by treating the item
as the vectors for which youre computing similarity, and the users
as the dimensions Allows for recommending similar items to a user
after theyve made a purchase Amazon uses a variant of this
algorithm This is an example of item-to-item collaborative
filtering Slide 9 Adding ROI to the Equation: an Example with Nave
Bayes When recommending products, some may generate more margin for
the firm than others Some algorithms can take cost into account
when making recommendations Nave Bayes is a commonly used
classifier that allows for the inclusion of marginal value of a
product sale in the recommendation decision Slide 10 Nave Bayes
Bayes theorem tells us the probability of our beliefs being true
given prior beliefs and evidence Nave Bayes is a classifier that
utilizes Bayes theorem (with simplifying assumptions) to generate a
probability of an instance belonging to a class Class likelihood
can be combined with expected payoff to generate the optimal payoff
from a recommendation Slide 11 Nave Bayes Continued How does the NB
algorithm generate class probabilities, and how can we use the
algorithmic output to maximize expected payoff? Lets say we want to
figure out which of two products to recommend to a customer Each
product generates a different amount of profit for our firm per
unit sold We know the target customers past purchasing behavior,
and we know the past purchasing behavior of twelve other customers
who have bought one of the two potential recommendation products
Lets represent our knowledge as a series of matrices and vectors
Slide 12 Nave Bayes Continued Slide 13 NB uses (independent)
probabilities of events to generate class probabilities Using Bayes
theorem (and ignoring the scaling constant) the probability of a
customer with past purchase history (a vector of past purchases)
buying item is: P ( 1, , i | j ) P ( j ) Where P ( j ) is the
frequency with which the item appears in the training data, and P (
1, , i | j ) is P ( i | j ) for all i items in the training data
That P ( 1, , i | j ) P ( j ) = P ( i | j ) P ( j ) is dependent up
on the assumption of conditional independence between past
purchases Slide 14 Nave Bayes Continued In our example, we can
calculate the following probabilities: Slide 15 Now that we can
calculate P ( 1, , i | j ) P ( j ) for all instances, lets figure
out the most likely boat purchase for Eric: These probabilities may
seem very low, but recall that we left out the scaling constant in
Bayes theorem since were only interested in the relative
probabilities of the two outcomes Nave Bayes Continued
P()ToysGamesCandyBooksBoat EricSquirt GunLifeSnickersHarry Potter?
Sailboat6/123/122/12 3/120.00086806 Speedboat6/121/122/121/12
0.00004823 Slide 16 So it seems like the sailboat is a slam dunk to
recommend. Its much more likely (18 times!) for Eric to buy than
the speedboat. But lets consider a scenario: lets say our
hypothetical firm generates $20 of profit whenever a customer buys
a speedboat, but only $1 when they buy a sailboat (outboard motors
are apparently very high margin) In that case, it would make more
sense to recommend the speedboat, because our expected payoff from
the speedboat recommendation would be 11% greater ($20/$1
*.0000048/.00087) than our expected payout from the sailboat
recommendation This logic can be applied to any number of products,
by multiplying the set of purchase probabilities by the set of
purchase payoffs, taking the maximum value as the recommended item
Nave Bayes Continued Slide 17 Challenges While recommendation
algorithms in many cases are relatively simple as machine learning
goes, there are a couple of difficult problems that all
recommenders must deal with: Cold start problem How do you make
recommendations to someone for whom you have very little or no
data? Data sparsity With millions of items for sale, most customers
have bought very few individual items Grey and Black sheep problem
Some people have very idiosyncratic taste, and making
recommendations to them is extremely difficult because they dont
behave like other customers Slide 18 Dealing With Cold Start
Typically only a problem in the very early stages of a user-system
interaction Requiring creation of a profile for new users can
mitigate the problem to a certain extent, by making early
recommendations contingent upon supplied personal data A
recommender system can also start out using item-item
recommendations based upon the first items a user buys, and
gradually change over to a person-person system as the system
learns the users taste Slide 19 Dealing With Data Sparsity Data
sparsity can be dealt with primarily by two methods: Data
imputation Latent factor methods Data imputation typically uses an
algorithm like cosine similarity to impute the rating of an item
based upon the ratings of similar users Latent factor methods
typically use some sort of matrix decomposition to reduce the rank
of the large, sparse matrix while simultaneously adding ratings for
unrated items based upon latent factors Slide 20 Dealing With Data
Sparsity Techniques like principal components analysis/singular
value decomposition allow for the creation of low rank
approximations to sparse matrices with relatively little loss of
information Slide 21 Dealing With Sheep of Varying Darkness To a
large extent, these cases are unavoidable Feedback on recommended
items post purchase, as well as the purchase rate of recommended
items, can be used to learn even very idiosyncratic preferences,
but take longer than for a normal user Grey and black sheep are
doubly troublesome because their odd tendencies can also weaken the
strength of your engine to make recommendations to the broad
population of white sheep Slide 22 References A good survey of
recommendation techniques Matrix factorization for use in
recommenders Article on the BellKor solution to the Netflix
challenge Article on Amazon's recommendation engine
