Data Mining: Bookstore Recommendation Project

Source: Slideplayer
I wrote a brief introduction about this project here, be sure to check it out. Today, I will be sharing the steps we took in solving this team project, the algorithm we used and finally the outcome of the project. I love to keep my articles short, simple and straight to the point so if there are few gaps or questions. You can drop them in the comment section or contact me. I will try to provide a solution if I can. Otherwise let's get right into it.

The task was to recommend set of books to promote together for a specific customer group. Also, to give recommendations to the CEO so that he/she can make the right decision with the aim to improve decision making and stay profitable. The good thing about this project is that the sample dataset was given to us so all we needed to do was work with the data and provide the necessary recommendations. 

WEKA was the recommended software we were asked to use to solve this problem. I will be breaking down the steps we used to solve this problem in WEKA  accompanied with screenshots of our outputs.

1) We launched WEKA application 
2) We imported the given dataset. Note: The dataset we were given was in ARFF format
3) The first thing to note in this kind of problem is that, it is a recommendation problem which obviously streamlined our decision on the algorithm to use.
4) We decided to use Association mining rule. We applied the Apriori Algorithm. Why? Simply because this rule/algorithm are known to be used for finding relationships between frequent itemsets, correlations and associations. 
5) After deciding on the rule that was applicable to our project, we needed to be sure there were no missing data in our dataset and also know the type of data that we were given which is called pre-processing stage.

A. Pre-processing Stage
1) Our data had 599 instances with 12 attributes. 
2) The data was initially a numeric (REAL) data type. However, in order to be able to apply Apriori algorithm, the data type was converted to nominal using 'NumericToNominal' unsupervised filter in WEKA.

Raw Data Visualization in WEKA
Data after Conversion in WEKA
3) The 'ID Transaction' attribute was removed simply because it does not add any value to the data mining approach. 
4) At the end of the pre-processing stage, the dataset consisted of  11 attributes with either 0s or 1s. Where 0 indicated that the item was not bought and 1 indicated that an item was bought.

B. Analysis Stage
1) In this stage, the Apriori algorithm was applied on the dataset. However, we discovered that WEKA built the model based on only unpurchased items πŸ˜• which was not our intention. Our aim is to give recommendations based on purchased items or 1s. How then we do we move forward from hereπŸ˜–
2) WEKA of course has a feature to solve this which was what we applied and voila we got some juicy outputs to work withπŸ˜‹. What then is this feature?
3) Well, in the Apriori algorithm settings there is a feature called "treatZeroAsMissing" which by default is set to "False" so we set this feature to "True" and yes πŸ’ͺ that was it.

4) We reran the algorithm but no best rules were found at the default 'minMetric' of 0.9, which indicated that no best rules were found at a 90% confidence.

5) However, we reduced the 'minMetric' to 0.8, 0.7 and even 0.6.and we were able to get some really good combinations which we used for our recommendation/ solution to the problem.

Output at 0.8 confidence level (3 best rules found)
Output at 0.7 confidence level (10 best rules found)

Output at 0.6 confidence level (10 best rules found)

C. Analysis of Results
1) Each of the rules that we found contained ‘A=>C’ which means that if a set of antecedents (A) are purchased, then there is a probability that Consequent (C) will also be purchased. For example, for the output at 0.8 confidence level, 78 transactions contained purchase of a Youthbook and a Cookbook (A). Out of those transactions, 67 instances contained a Childbook (C). The latter part is referred to as ‘Support’ for the Consequent. The Confidence score shows how confident the association rule is, given the dataset. It is calculated as: C/A: 67 / 78 = 0.86.

2) Another interesting parameter is the ‘Lift’ which is defined as how likely it is to have all antecedents and consequent in one single transaction in comparison to the entire transaction dataset. Basically the larger the lift ratio, the more significant the association of the itemset. In order to calculate Lift, first we needed to figure out the ‘Expected Confidence’ which is the probability of the purchase of consequent regardless of the antecedents. As an example, looking at the first rule (0.8 confidence), the total number of transactions containing Childbook (250) divided by the total number of transactions (599): 250/ 599 = 0.417362270 (approximately 0.42)

3) After calculating the Expected Confidence, the Lift can then be calculated. This is the ratio of the confidence and the expected confidence: 0.86 / 0.417362270 = 2.06

4)With the confidence score of 86% and the lift score of 2.06, this rule can be considered as a strong association. That is just the analysis of one rule. I wouldn't be going through all the analysis of all the rules in this article.

5) After building the model, these three best rules were found by Weka:
  • If a Youthbook and a Cookbook are purchased in one transaction, there is 86% confidence that a Childbook will be purchased
  • If a Cookbook and a Refbook are purchased in one transaction, there is 83% confidence that Childbook will be purchased
  • If a Cookbook and a Geogbook are purchased in one transaction, there is 82% confidence that Childbook will be purchase

D. Our Recommendation
Based on our analysis and the results from WEKA, the decision/business model that we would recommend is that: since Childbook has a relatively high correlation with Youthbooks, Cookbooks, Refbooks and GeogBooks, then they can be promoted together.

In conclusion, when trying to solve a data mining problem. There are several ways to go about it. There are also several ways to interpret your results after your analysis. However, I would recommend that you understand the basis behind WEKA if you are not familiar with it. This will give you a better understanding of whatever project you are given and diferent ways to go about it.

I hope this article helps someone out there trying to get a hang of a similar project. Below are the list of some useful links that were very useful for us while we were solving this problem.
1) Building a market basket model
2) Market Basket Analysis with Association Rule Learning
3) Lift in Association Rule 

Hey before you go, if you like this article, consider buying me a coffee by clicking here. Until next time...πŸ’‹