Look-alike Model (LAM) - Technical Explanation for LAM Audiences for Flash Studies

When creating audience segments for your Flash Study - Resonate uses the LAM methodology to ultimately generate more leads for clients and increase their conversions. The LAM methodology is an instance-based Machine Learning approach. The algorithm discovers similarities between internet users based upon their behavior online. Currently the model uses the domain level URLs that the users visited during the two weeks prior to modeling. The collected behavior for all users usually amounts to tens of thousands of URLs.

The method begins with a dimensionality reduction step through Singular Value Decomposition (SVD). SVD can handle enormous matrices (in our case the original data matrix A has about 250 million users by approximately 60 thousand behavioral signals). The original data matrix A is very sparse and does not provide an adequate answer to how relevant two users are to each other.

After several passes over the data, the SVD returns three matrices which when multiplied together returns approximately the original matrix A. One of the matrices SVD produces: U has columns that form an orthonormal basis for the users’ space. The decomposition is parametrized with a number k which indicates how many features to keep around (in our case k= 300). A key insight of SVD is that only a small number of features is important to represent users’ behavior. We can produce a relevance score between two users by computing the cosine similarity between their two-column vectors in U. We think of vectors that point in the same direction (high value of cosine) in the high-dimensional users’ space as relevant (similar) to each other.

Further details:

Rationale

We found this methodology to be optimal in targeting settings where standard KPIs apply as well as scenarios where sample sizes are small and/or where defining a null set is intractable (e.g., single-class classification). For example, did a person not click on an advertisement because creative is ill-suited or because it was not even presented?

The heuristic

We calculate the “dot-product” for survey respondents’ vectors with all online users’
vectors. When vectors are normalized the dot-product is the same as the cosine between the vectors.

• The dot product is the sum of the products of each component of the two vectors.

• The dot product is a value expressing the angular relationship between two vectors.

• It is thus a judgment of orientation (not magnitude).

• Angle = 0 -> cosine = 1.0: vectors are codirectional; or maximally similar

• Angle = 90 -> cosine = 0.0: vectors are maximally dissimilar

Scale and Relationships

This methodology is unusual in that there are almost no gradient-tunable parameters and due to the nature of the heuristic ranking, requires apriori user input to determine the size of the delivered segments or relationships between segments. Currently, we support scale and relationships to be determined apriori with the following settings:

1. Cosine threshold – return all devices with cosines exceeding a specific threshold

2. Top N – sort predictions and return the highest N devices

3. Proportional – save the highest N% of devices

4. Boolean relationships that must be maintained: When there are complicated relationships (e.g., exclusivity and subsets) between segments we first define the Boolean logic to capture segments prior to any modeling and then apply that logic to the final ranking.

Performance

We can assess the performance of this method in multiple ways.

1. Targeting – a series of cross-vertical experiments indicate this methodology is our best tool for maximizing standard targeting KPIs.
2. Recall of initial segments – If the size of the inputs exceeds 1000 samples, we organically produce a holdout set that could be used to compute the recall of holdouts in the final delivery. In practice, however, the methodology operates between a tension of theoretical expectations and small sample sizes and recall of the training set is the only performance metric we can track.