The Prediction Lover’s Handbook
Assessment tools for better-informing decisions have proliferated. Here’s an insider’s guide to prediction and recommendation techniques and technologies.
Predictive analytic technologies incorporate statistical, rule-based and/or social networking techniques that can either be used to recommend actions for buyers or to predict customer behavior for business objectives. Predictive applications seek to discover and apply patterns in data to predict the behavior of customers, products, services, market dynamics and other critical business transactions. Recommendation engines (often called “recommenders”) attempt to influence or predict what customers want, enjoy or need. They also often incorporate a company’s business objectives and available offerings into the recommendations.
Prediction and recommendation applications use one or more techniques to improve the accuracy of their predictions. Our prediction: These techniques will continue to improve and gain broad market acceptance.
Biological responses analysis
Used by: Developers of successful television series for children, including “Sesame Street” and “Blue’s Clues”; Boston-based Innerscope Research for assessing responses to advertising and television content.
Strengths: Takes guesswork out of questions involving human response to cultural products.
Weaknesses: Highly invasive. Difficult to understand reasons behind a particular biological response.
Prospects: Will probably grow in popularity with increased understanding of neurological processes despite ethical concerns.
Used by: Platinum Blue Music Intelligence to group sound attributes that are attractive to particular customers; marketing departments in retail and consumer product companies to determine strategies for cross-selling products and to identify new customer segments and product needs.
Strengths: Fast and scalable. Usable by companies lacking detailed consumer behavioral data.
Weaknesses: Lacks detailed personalization possible with other techniques.
Prospects: Useful for analyzing customer behaviors for cultural products, but less so for recommendation and real-time prediction.
Attributized Bayesian analysis
Used by: TiVo Inc. (for new customers or television shows); Choicestream Inc., of Cambridge, Massachusetts, for recommending a variety of cultural products; also used in textual analysis/text mining applications.
Strengths: Typically offers more refined recommendations than collaborative filtering. Useful for “cold start” applications where little personalized data is available. Analysis can be quickly developed, and may produce surprising choices.
Weaknesses: Difficult to determine and classify attributes. Users often depend on databases of offering attributes prepared by others; if no such database exists, classifying attributes can be daunting.
Prospects: Will be adopted where there exist attribute databases upon which to draw.
Content-based filtering/Decision Trees
Used by: Web sites like CNET to help customers choose among several products (what cell phone should I buy?); game vendors to incorporate decision trees allowing individuals to personalize a video game; call center personnel to determine script for cross-selling products to customers; investment managers balancing portfolios.
Strengths: Easy to interpret by subject experts. Useful even when data is limited. Rules are transparent and manageable by subject experts.
Weaknesses: Unstable since small variations in the model can result in large variations in responses. Not suitable for highly complex applications where the trees become too complex to interpret.
Prospects: Useful in principle to calculate and portray probabilities of multiple cultural product purchases, but unlikely to be widely used.
patterns in the preferences and purchasing behavior of a buyer to those of other buyers. Item-to-item filtering compares an item selected by the buyer with all other items in the database — for example, 90 percent of people who bought A also bought B. User-to-user filtering places buyers in a community of people with similar purchasing behavior, and makes recommendations based on what others buy or prefer.
Used by: Amazon; TiVo Suggestions; Digg; Delicious; Last.fm’s Audioscrobbler; Netflix; Yahoo!; Pandora Media; Apple’s Genius; Microsoft MixView for predictive analysis and generating recommendations.
Strengths: Currently the most widely used predictive technology.
Weaknesses: Inappropriate for new or unknown products. Biased toward high-volume products. Also, does not distinguish between items bought for one’s own use or purchased for others.
Prospects: The fastest-growing tool for cultural product recommendation, but needs further refinement to personalize recommendations. Search engines such as Google Inc. will add CF to recommend Web sites.
Neural network analysis
Used by: Financial services companies to predict creditworthiness and detect fraud; London-based Epagogix Ltd. to predict successful movies from script attributes.
Strengths: Good for generating predictions in applications with complex, “dirty” data. Can identify previously unrecognized patterns.
Weaknesses: Reasons for recommendations are often unclear. Doesn’t always scale well. Requires large quantity of data and considerable expertise to develop.
Prospects: In an era of greater corporate transparency and explainability, neural nets will remain limited to very specialized applications.
Prediction (or opinion) markets
Used by: Hollywood Stock Exchange; Interactive Music Exchange for Fuse Networks.
Strengths: Easy to set up. Becomes more accurate with larger samples and more knowledgeable, frequent contributors.
Weaknesses: Crowd perceptions are only useful when they have enough data to form opinions and are motivated to share them. Paying real money to successful bettors is illegal in the United States. Participant dropout rate is high.
Prospects: Prediction markets are growing in general, but it’s too soon to tell for cultural product applications. Financial institutions, technology companies and retailers will experiment with the technique to help them predict shifts in consumer tastes, market and liquidity in turbulent times.
Social network-based recommendations
Used by: MySpace Music; MTV SoundTrack; Last.fm.
Strengths: Can leverage existing social networking sites like Facebook Inc. and MySpace.com. Well suited to new and small businesses lacking a customer base to identify and reach potential customers.
Weaknesses: Recommendations are not scientifically determined. Social network relationships can be superficial and have less impact than other techniques. Large corporations may have better ways to reach potential customers.
Prospects: Will generally be augmented by other recommendation approaches that rely on data analysis.
Used by: Google Inc. and other search engines; marketers analyzing consumer and opinion-maker sentiment on the Web; also used by music and other Web sites recommending products/services.
Strengths: Analyzes unstructured, nonquantitative written material. Enables fast response to unfavorable postings by spotting them quickly.
Weaknesses: Automated interpretation of the semantic “meaning” of text is still a developing field. Can lead to considerable “noise” in the data.
Prospects: Explosive growth as researchers develop better techniques to mine and interpret text.
Used by: Movie studios, music companies and publishers to predict how many copies of a particular product to manufacture and distribute based on past results or early indicators.
Strengths: Good for generating predictions when there is considerable data from the past. Widely understood.
Predictions are often accurate when prediction period is short range.
Weaknesses: Only as good
as the data, independent variables and hypotheses of analysts.
Prospects: “Workhorse” technique that will continue to be useful in its application space.