Lift during the MB inspect

[DanielH]

Hello all,

I have been trying my hand at the predictive tooling from Alteryx and used the MB Rules - MB Inspect to do a market basket analysis. The results are quite good with smaller datasets and I can understand most of the details. However the "Lift" still confuses me a little bit. Alteryx help says the following:

"Lift is the proportion of transactions that contain the RHS and LHS items relative to the proportion of time we would expect this to occur if the items were independent (unrelated) to one another. If the lift measure is one, then the RHS and LHS item(s) as often in transaction as they are expected to when the items have no relationship to one another. If the measure is less than one, then they occur less frequently together than would be expected if they were unrelated to one another (they "repulse" one another). If the measure is greater than one, then the LHS and RHS item(s) occur together more often then is expected by random chance. In general, we are interested in association rules that have a lift substantially greater then one. However, as this minimum lift criteria increases, we have fewer returned rules. A value of zero "turns-off" this filter, which is only applicable for association rules."

This might not be the most technical question but what does a lift of 13.84 or 50 mean? It's the one thing I do not understand yet about this analysis.

Kind regards,

Daniel

[chris_love]

Hi Daniel

Perhaps this diagram will help.

Chris

[DanielH]

Hello Chris,

Thank you for your reply. Are these scores independent from one another? (e.g. a score of 80 at one pair is not the same as a score of 80 at another pair). And can you say anything about the individual scores other than that they are "almost always" bought together? Im hoping to calculate a projected profit when pairing certain items.

Daniel

[chris_love · Accepted solution]

Good question. Let me revise my explanation.

The lift score actually tells you how like they are that the fact they are bought together is not random, the higher the score the less likely it is to be random. There are several factors here:

1. Total number of purchases in whole dataset

2. Total purchases of item 1

3. Total purchases of item 2

4 Total purchases of items 1 and 2 together

Clearly if there are only 10 overall items purchased but item 1 and 2 were bought together twice then we don't have enough information to say the interactions aren't random so you may get a score over 1 but not much. A very high score might indicate that there were a lot of purchases but very few of items 1 and 2, but when they were bought they were bought together. or it may indicate that there were a lot of purchases of 1 and 2 and a high frequency within that of 1 and 2 being bought together.

As you can see its not easy to say this score means X.

If you want to highlight the effectiveness of promotions / combinations of items and choose new ones then one way of doing it might be:

a. look at previous promotions, look before and after and look at the lift before and after. did the promotion actually increase the chance that items were bought together? i.e. did the lift increase.

b. use that to work out which combinations work and how much the promotion uplifted profit. 

c. use those product combinations to predict which other cominations may work (e.g. by using a nearest neighbour score to find similar products)

d. interpolate the projected profit from the historic promotions to the sales for these new promotions.

Make sense?

[DanielH]

Thanks for your reply!

It makes sense now. I don't have to search for the specific meaning of lift anymore and can advise on certain itemsets or promotions we need to try. My idea is to split the dataset or add an identifier indication a promotion or sales period and see what kind of impact lift has when combining frequent itemsets.

Kind regards,

Daniel