Category: Data Science

Tanner Thompson describes usage of a multi-armed bandit algorithm to drive conversions:

The functional idea behind a bandit algorithm is that you make an informed decision every time you assign a visitor to a test arm. Several bandit-type algorithms have been proved to be mathematically optimal; that is, they obtain the maximum future revenue given the data they have at any given point. Gittins indexing is perhaps the foremost of these algorithms. However, the trade-off of these methods is that they tend to be very computationally intensive.

This article doesn’t show any code, but it is useful for thinking about the problem.

Anand Ramanathan has a strong critique of “data science” as it stands today:

Illustration: Consider the sentiment-tagging task again. A Q1 resource uses an off-the-shelf model for movie reviews, and applies it to a new task (say, tweets about a customer service organization). Business is so blinded by spectacular charts [14] and anecdotal correlations (“Look at that spiteful tweet from a celebrity … so that’s why the sentiment is negative!”), that even questions about predictive accuracy are rarely asked until a few months down the road when the model is obviously floundering. Then too, there is rarely anyone to challenge the assumptions, biases and confidence intervals (Does the language in the tweets match the movie reviews? Do we have enough training data? Does the importance of tweets change over time?).

Overheard: “Survival analysis? Never heard of it … Wait … There is an R package for that!”

This is a really interesting article and I recommend reading it.

Bernard Antwi Adabankah uses a decision tree to model price changes:

The sample included N = 262 individual orders for Interlocking Hearts Design Cake Knife/Server set with OrderItemSKU as 2401 from the period ranging from 1^st March 2014 to 20^th April 2016 with an ecommerce company which sells on Amazon.co.uk

The Profit response variable is measured as the product sale price on amazon.co.uk which includes amazon.co.uk commission and any applicable postage costs less the purchase price of the Hearts Design Cake Knife/Server set from the supplier.

Read the comments for a couple good critiques of the article.

The folks at ELEKS discuss what to do when a single time series model just won’t cut it:

With the emergence of the powerful forecasting methods based on Machine Learning, future predictions have become more accurate. In general, forecasting techniques can be grouped into two categories: qualitative and quantitative. Qualitative forecasts are applied when there is no data available and prediction is based only on expert judgement. Quantitative forecasts are based on time series modeling. This kind of models uses historical data and is especially efficient in forecasting some events that occur over periods of time: for example prices, sales figures, volume of production etc.

The existing models for time series prediction include the ARIMA models that are mainly used to model time series data without directly handling seasonality; VAR models, Holt-Winters seasonal methods, TAR modelsand other. Unfortunately, these algorithms may fail to deliver the required level of the prediction accuracy, as they can involve raw data that might be incomplete, inconsistent or contain some errors. As quality decisions are based only on quality data, it is crucial to perform preprocessing to prepare entry information for further processing.

Treating time series data as a set of waveform functions can generate some very interesting results.

Christina Prevalsky makes the case for data curation:

The gaining popularity of self-service analytical tools such as Tableau increases the necessity of having curated data in your database. These tools aim to allow the end users to intuitively query data “at the speed of thought” from the data warehouse and visualize the results quickly. That type of capability allows users to go through several different iterations of the data to really explore the data and generate unique insights. These tools do not work well when the underlying database tables have not been curated properly.

This is a difficult and lengthy process, but it’s vital; data minus context is a lot less relevant than you’d hope.

Peter Coates provides an extremely fast estimate of Levenshtein Distance:

If your application requires a precise LD value, this heuristic isn’t for you, but the estimates are typically within about 0.05 of the true distance, which is more than enough accuracy for such tasks as:

• Confirming suspected near-duplication.

• Estimating how much two document vary.

• Filtering through large numbers of documents to look for a near-match to some substantial block of text.

The estimation process is pretty interesting.  Worth a read.