In this paper I chose a specific question about writing. The core product data anomalies are common problems that occur at work, the Internet is a common interview questions. Here I combine online and sharing some of their own experiences, summarizes some of the thinking and analysis framework, so that we have a clear focus in the face of such problems.
About Case
A flow of information APP Nikkatsu usually stable between 79w-80w, but since June 13 suddenly fell 78.8w, to June 15 has dropped 78.5w, then the product owner anxious, so you as soon as possible investigation and data about the reasons for decline. Such a problem for most people, it is quite troublesome, because for 80w order products, a twenty thousand is not a very big fluctuations, but the reason is that you are troubleshooting.
Get this, will feel not know where to begin start-point analysis? It does not matter, we have common routines stroke clear, and then look back on this case.
Core point: first hypothetical data error cause, after data verification hypothesis.
We do not recommend the first step toward their own data to dismantle, Nikkatsu factors influence a lot of data, it is impossible to compare all of the dimensions dismantling one by one, easy to waste time but did not find anything of value. The core reason for the exception for data analysis is combined with previous experience and a variety of information, identify the most likely cause hypothesis, conduct multi-dimensional analysis to verify the hypothesis by splitting the data, locate the problem. Process may assume or create a new adjustment in the original assumptions on the original hypothesis until you locate the cause.
The first step: to confirm the authenticity of data
Before embarking on the analysis began, it is recommended to confirm the authenticity of the data. We often encounter data services, data reporting, BUG on statistics, there will be outliers in the data report. So, look for the data flow related products and research and development to confirm the authenticity of the data under the bar.
Step two: According to several common dimensions of preliminary data split
The impact factor calculation: Each piece of data should be normal as ever to do comparison, calculate the impact factor.
Effect factor = (the amount of today - yesterday amount) / (total today - yesterday's total)
The greater the impact factor, indicating here is the main point drop
These are several common initial dimension resolved by splitting the initial positioning approximate range reasons.
The third step: positioning abnormal range, further assumptions are made
For the initial positioning of the affected area for further investigation. In three dimensions do assume, is recommended for the data to build a group dedicated to unusual problems, one way or another appropriate product, technical, operational staff together, understanding data anomalies near the point in time did what products, operations, technology side adjustment.
Considering the previous data anomalies reason, technical product operations side to adjust the initial positioning of influence is most likely caused by what, combined with their business experience to determine the most likely cause of several assumptions, these assumptions give priority row data validation, one by one investigation.
Finally: segmentation hypothesis, to establish the cause
In addition to the above-described dimension may be too many segmentation analysis, in that said core logic after a verified hypothesis, based on this assumption is true, the data splitting finer dimensions. We need to remember that this analysis way, when some cause data anomalies guess is that as long as the antithesis of reason to find segments represented do comparison, we can prove or disprove our guess, until finally find the real reason.
case study
These are the core data analysis routines unusual, is not it just do not know where to start to get the problem analysis, now I feel that there are many points can go about? Let's go back just a case of it. According to the routine, first we split the old and new users active amount, below (the old user LHS, RHS new user):
Users found the old days live more stable, but new users since June 13 serious decline, then calculate the new and old customers influence coefficient:
Old user influence coefficient = (77.89-78) / (78.8-79.5) = 0.16
New user influence coefficient = (0.98) / (78.8-79.5) = 0.84
New user influence coefficient 0.84, indicating DAU decline in new user who is a clearly a backward range segment, the new user consists of what?
New user channel = channel 1 + 2 + 3 + channels, other channels , so we put a new user Nikkatsu split by channel:
By splitting channels, we found the channel 3 from June 13 onwards serious decline in new users, so we locate the problem in the channel 3, channel 3 should be the channel effect of a problem. 3 channels of locating the person in charge with specific reasons, to reduce the amount of channels clue? Channel conversion decreases? Channel platform of the problems? After find out why, and then for reasons solve problems, develop channel optimization strategy.
Last but not least
So far this article has come to an end, a detailed description of the core abnormalities of routine data analysis as well as an easy to tell everyone to understand the small case, I believe we next time you encounter such problems, there is at least one clear place to start.
Some say to you is: In order to facilitate understanding of the data this small case is my imaginary, problems positioning process is relatively simple. However, in actual business, the data may be due to abnormalities affecting many (Benpian only talked about some internal factors, external environment and competition will also affect the core data on fact), sometimes also need to build a statistical model to do Some quantitative analysis.
It may take a few days to continue to troubleshoot the problem, this process is tedious and boring, assuming that there may be frustration validation fails, perhaps busy for a long time but in the end did not find out why.
In fact, this is a normal thing, abnormal data analysis problems even for a senior data analyst is a headache. So we need to pay more attention to data changes in routine work, along with sensitivity and enhance the business of familiar data analysis for data anomalies we will be more skilled, more quickly find the problem.
In this part we hope to have a real help, want to know more follow-up data analysis related Internet content, welcome attention thumbs forward, welcome to discuss more topics.
Attach frame Figure own summary:
- The first step: to confirm the authenticity of the data
before embarking on analysis begins, it is recommended to confirm the authenticity of the data. We often encounter data services, data reporting, BUG on statistics, there will be outliers in the data report. So, look for the data flow related products and research and development to confirm the authenticity of the data under the bar. - The second step: a clear definition of
detailed analysis by clearly defined indicators - The third step: to locate the problem
preliminary splitting dimension, locate the approximate range of reasons. The impact factor calculation: Each piece of data should be normal as ever to do comparison, calculate the impact factor. Influence coefficient = (amount today - yesterday's volume) / (total today - yesterday the total amount) the greater the impact factor, indicating here is the main point drop- Demolition of users: the user type (old and new, premium renewals, new, reflux, etc.), user portrait in all aspects (regions, channels, age, gender, occupation, etc.)
- Demolition platforms: IOS / Android
- Demolition Version: new / old version
- Demolition area: Province / State
- Split time: access period, seasonal, cyclical product (YoY)
- Login demolition channels: app / applet / PC / m terminal, etc.
- Demolition of entry: Point icon to enter / push invoking etc.
- Step four: indicators dismantling
dismantling indicators, further positioning reasons- Dismantling formula
- Points out the limitations of dismantling method: hypothesis defects, defect distribution, estimate a conservative / aggressive
- Step Five: Analysis
of dimensions, metrics broken down cause analysis, considering past data anomalies reason, technical product operations side to adjust the initial positioning of influence is most likely caused by what, combined with their business experience to determine several the most likely reason is assumed that, to discharge these assumptions priority data validation, one investigation.- external factors
- PEST method
- Product Research Methodology
- market
- Competing products
- Society: public life, values, consumer psychology
- Internal factors
- Product side:
- Adjustment function
- Policy adjustment
- Technical side:
- Interface unstable
- system error
- Web page does not open, slow to load
- Operating side:
- Operations Strategy
- Advertising
- Operating activities (mention active, promote retention, pull Paid)
- push effect
- Pull new channels
- Product side:
- external factors
Just need to enjoy music, high-frequency
Impact of pricing: cost, demand, competition
