Method and apparatus for recommendation engine using pair-wise co-occurrence consistency

• US 20070094066A1
• Filed: 01/06/2006
• Published: 04/26/2007
• Est. Priority Date: 10/21/2005
• Status: Active Grant

• Current Assignee: Fair Isaac Corporation
• Original Assignee: Fair Isaac Corporation
• Inventors: Momma, Michinari, Kumar, Shailesh, Chow, Edmond
• Granted Patent: US 7,801,843 B2
• Time in Patent Office: Days
• US Class Current: 706/61
• CPC Class Codes: G06Q 10/04 Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem" market predictions or forecasting for commercial activities G06Q30/0202
• CPC Class Codes: G06Q 10/06 Resources, workflows, human or project management; Enterprise or organisation planning; Enterprise or organisation modelling
• CPC Class Codes: G06Q 30/02 Marketing; Price estimation or determination; Fundraising
• CPC Class Codes: G06Q 30/0201 Market modelling; Market analysis; Collecting market data
• CPC Class Codes: G06Q 30/0202 Market predictions or forecasting for commercial activities
• CPC Class Codes: G06Q 30/0204 Market segmentation
• CPC Class Codes: G06Q 30/0224 based on user history
• CPC Class Codes: G06Q 30/06 Buying, selling or leasing transactions
• CPC Class Codes: G06Q 30/0605 Supply or demand aggregation

Abstract

The invention, referred to herein as PeaCoCk, uses a unique blend of technologies from statistics, information theory, and graph theory to quantify and discover patterns in relationships between entities, such as products and customers, as evidenced by purchase behavior. In contrast to traditional purchase-frequency based market basket analysis techniques, such as association rules which mostly generate obvious and spurious associations, PeaCoCk employs information-theoretic notions of consistency and similarity, which allows robust statistical analysis of the true, statistically significant, and logical associations between products. Therefore, PeaCoCk lends itself to reliable, robust predictive analytics based on purchase-behavior.

Claims

1. A customer centric, data driven, transaction oriented purchase behavior apparatus, comprising:

a recommendation engine for processing multidimensional information comprising raw customer transaction history for customers to whom one or more products are recommended, a set of products in a recommendation pool that may be recommended to customers, and a set of times at which recommendations of said products have to be made to said customers; and

said recommendation engine generating a propensity score matrix with a score for each combination of customer, product, and time to allow merchants to match their content with customer intent;

wherein said recommendation engine offers the right product to the right customer at the right time at the right price through the right channel so as to maximize the propensity that the customer actually takes-up the offer and buys the product.

2. The apparatus of claim 1, further comprising:

one or more business constraints for filtering or customizing any of said customer, product and time information.

3. The apparatus of claim 1, further comprising:

means for capturing customer'"'"'s purchase behavior in the form of graphs.

4. The apparatus of claim 1, further comprising:

a post-processor, wherein said recommendation engine uses only customer history to create propensity scores that capture potential customer intent, said post-processor allowing retailers to adjust scores to reflect business objectives.

5. The apparatus of claim 1, further comprising:

a business rules engine for business constraints and objectives that are implemented as business rules, wherein said rules are implemented at a stage where propensity scores are used for any of creating top product recommendations per customer for a given time creating top customer recommendations per product for a given time, or finding a best time to recommend a particular product to a particular customer.

6. The apparatus of claim 1, further comprising:

a channel specific deployment mechanism by which decisions about a right channel which depend on the nature of a product being recommended and a customer'"'"'s channel preferences are made by a merchant to deliver those recommendations using various channels, once recommendations are created for each customer.

7. The apparatus of claim 1, further comprising a plurality of recommendation modes, comprising any of:

product-centric recommendations;

customer-centric recommendations; and

time centric recommendations.

8. The apparatus of claim 1, further comprising a plurality of recommendation triggers, comprising any of:

real-time or near-real time triggers which require that recommendation scores are updated based on real-time triggers; and

batch-mode triggers which require that recommendation scores are updated based on pre-planned campaigns, wherein a batch process may be used to generate and optimize said campaigns based on recent customer history.

9. The apparatus of claim 1, wherein said propensity scores depend on customer history, comprising any of:

a current purchase for anonymous customers, where a customer history is not available;

recent purchases, where customer history is known, and wherein purchase behavior is highly time-localized;

entire history as a market basket, where a time component is not as important and only what customers bought in the past is important, and wherein an entire customer history weighted by product recency is used while ignoring a time component;

entire history as a sequence of market baskets, where a time interval between successive purchases of specific products is important, and wherein customer history is treated as a time-stamped sequence of market baskets to create precise and timely future recommendations; and

products browsed, where a customer browses a product to consider purchasing, and wherein the fact that a customer took time at least to browse these products shows that he has some interest in them and, therefore, even if he does not purchase them, they are still used as part of a customer history along with products he did purchase.

10. A method for recommendation scoring, comprising the step of:

creating a propensity or likelihood score for what a customer might buy in the near or far away future based on customer history.

11. An apparatus for recommendation scoring, comprising:

a recommendation engine, comprising any of;

a market basket recommendation engine (MBRE) that treats customer history as a market basket comprising products purchased in the recent past; and

a purchase sequence recommendation engine (PSRE) that treats customer history as a time-stamped sequence of market baskets.

12. The apparatus of claim 11, said MBRE comprising:

means for generating an MBRE recommendation score comprising a propensity score which is computed via any of;

a Gibb'"'"'s aggregated consistency score which aggregates consistencies between products in a market basket with a target product;

wherein aggregate weighting depends either on consistency between products in the market basket with the target products or on seedness of the products in the market basket assuming a single bundle or a mixture-of-bundles;

a single bundle normalized score, in which transaction data comprise a mixture of projections of multiple intentions, where a market basket represents a single intention and is treated as an incomplete intention and where adding a target product makes it more complete, wherein a propensity score is defined as the degree by which bundleness increases when a product is added; and

a mixture-of-bundles normalized score which first finds all individual bundles in a market basket and then uses a bundle that maximizes the single bundle normalized score, wherein said bundles are also compared against single products as well as a entire market basket.

13. The apparatus of claim 12, further comprising:

a purchase sequence context which uses a time-lag between any past purchase and a time of recommendation to create recommendations.

14. The apparatus of claim 11, said PSRE comprising:

means for generating a PSRE recommendation score comprising a propensity score which is computed via any of;

a Gibb'"'"'s aggregated consistency score which aggregates consistencies between products in a market basket with a target product;

wherein aggregate weighting depends either on consistency between products in the market basket with the target products or on seedness of the products in the market basket assuming a single phrase or mixture-of-phrases;

a single-phrase normalized score, where a purchase history represents a single intention that is treated as an incomplete intention and where adding a target product at decision time makes the intention more complete, wherein a propensity score comprises a degree by which phraseness increases when a product is added at decision time; and

a mixture-of-phrases normalized score, which first finds all individual phrases in a purchase sequence and then uses a phrase that maximizes a single phrase normalized score, wherein said score is also compared against all single element phrases, as well as an entire phrase.

15. The apparatus of claim 11, further comprising:

means for making post-processing recommendation propensity score adjustments, said adjustments comprising any of;

a first order, transaction data driven score adjustments, in which adjustment coefficients are computed directly from transaction data; and

a second order consistency matrix driven score adjustment, in which adjustment coefficients are computed from consistency matrices.

16. The apparatus of claim 11, further comprising:

means for making score adjustments, said adjustments comprising any of;

a first order seasonality adjustment for adjusting recommendation scores, where products that have a higher likelihood of being purchased in a particular season are pushed up in a recommendation list in a systematic way;

wherein said first order seasonality adjustment is made by first computing a seasonality score for each product, for each season, wherein said score is high if a product is sold in a particular season more than expected; and

wherein said seasonality score is created as follows;

define seasons by a set of time zones; and

then compute a seasonal value of a product in each season, as well as its expected value across all seasons;

a first order value adjustment for creating a value-score for each product, wherein said value-score is normalized and combined with a recommendation score to increase or decrease an overall score of a high/low value product;

a first order loyalty adjustment for increasing customer loyalty and repeat visits by recommending to a customer those products that said customer bought in a recent past and by encouraging more purchases of said products;

wherein said first order loyalty adjustment is determined as follows;

create a value-distribution of each product that said customer purchased in the past;

compare said value-distribution to a value-distribution of an average customer or an average value distribution of that product; and

if a customer showed higher value than average on a particular product, then increasing the loyalty-score for that product for that customer;

a second order density adjustment for adjusting a recommendation score to push high density products for increased market basket sizes;

wherein said second order density adjustment for each product is determined by performing a Gibb'"'"'s aggregate of values which may include any of frequency, revenue, or margin of neighboring products, and using consistency between the product and its neighbors to compute, a weighting for Gibb'"'"'s aggregation.

a second order diversity adjustment for increasing diversity of a customer'"'"'s market basket along different categories or departments via a diversity score that is used in post-processing;

wherein said second order diversity score for each product is determined by deviation of the Gibb'"'"'s aggregated density scores across different categories or departments of products in the neighborhood of the said product; and

a second order life-time value adjustment that is computed from a purchase sequence context to push products that increase a life-time value of customers.

17. A method for quantifying and discovering patterns in relationships between entities comprising products and customers, as evidenced by purchase behavior, the method comprising the steps of:

applying consistency and similarity functions to transaction data generated by said entities;

performing a statistical analysis of statistically significant and logical associations between products;

analyzing product associations in a plurality of contexts comprising any of individual market baskets, a next visit market basket, or across all purchases in an interval of time;

wherein different kinds of purchase behavior associated with different types of products and different types of customer segments are revealed; and

combining multiple consistency matrices as long as they are at a same product level and are created with same context parameters, for either;

dealing with sparsity by using a smoothing model, where counts from an overall co-occurrence counts matrix based on all customers are combined linearly with counts of a segment'"'"'s co-occurrence matrix, resulting in statistically significant counts; and

creating interpolated solutions to either compare a particular segment against an overall population to find out what is unique in a segment'"'"'s co-occurrence behavior, or to interpolate between a segment and the overall population to create more insights and improve the accuracy of said recommendation engine.

18. The method of claim 17, further comprising the step of:

combining segment level and overall population level analysis by any of the following;

a counts combination, where raw co-occurrence counts from all customers, averaged per customer, is linearly combined with raw co-occurrence counts from a customer segment;

a consistency combination, where consistency measures of co-occurrence consistency matrices are combined; and

recommendation scores, computed for a customer based on an overall recommendation model, as well as a recommendation model based on a customer'"'"'s segment based recommendation model;

wherein two scores may be combined to come up with potentially more accurate propensity scores.

19. The method of claim 17, wherein the number of context instances for a given context type and context parameters depends on a number of customers, a number of transactions per customer, and a number of products per transaction

20. The method of claim 17, further comprising the step of:

addressing sparsity via any of;

product level count smoothing which uses a hierarchy structure of a product space to smooth out co-occurrence counts;

wherein for any two products at a certain product resolution, if either a margin or co-occurrence counts are low, then counts from a coarser product level are used to smooth counts at this level;

wherein said smoothing uses both a parent level and a grand-parent level if there is a need; and

wherein as statistical significance at a desired product level increases due to additional transaction data becoming available over a period of time, a contribution of coarser levels decreases systematically;

customization level smoothing, where co-occurrence counts or consistencies based on all customers are used to smooth the counts or consistencies of a segment;

wherein if there is a multi-level customer hierarchy with segments and sub-segments, then this approach is generalized to use a parent segment of a sub-segment to smooth segment counts; and

context coarseness smoothing, wherein context is chosen at a level of coarseness that results in statistical significance of the counts and consistencies;

wherein any combination of the foregoing techniques may be used, depending on the nature, quantity, and quality (noise-to-signal ratio) of transaction data.