Sunday, 11 February 2018

Updated* Machine Learning | Insights from Visualisation | Multi-dimensional data

Data sets for machine learning are multi dimensional. The number of dimensions depends on the data domain. For example data for a collaborative item-to-item recommender includes a user and an item. This can easily be fit to a co-occurrence matrix. However, data from medical diagnosis have more than 3 dimensions. This is the same for air pollution measurements, vehicle stats and other data sets. When the number of dimensions is 5 or less, it is easy to visualise data before deciding on an approach to machine learning.

Visualising data before applying machine learning has advantages.

1) Visually identify unique patterns in dimensions. Those dimensions are likely more important to help in decision making.

2) Help to decide candidate model(s). For example, would it be better to use a Random Forest classifiers or Support Vector Machine or Nearest Neighbour Classifier?

3) Select a subset of dimensions (before model fitting) that are more suitable for machine learning. Not all collected dimensions turn out useful. Subsequently adding dimensions leads to different use cases and insights.

The rest of the post is divided into
1) Visualization
2) Prediction results using Classifiers
3) Plotting feature weights from fitted models to confirm insights from visualization.

1) Visualization

The sample data set for this post is the Breast Cancer Wisconsin (Diagnostic) Data Set from University of California, Irvine Machine Learning Repository. The dimensions are measurements on cell nuclei. There are 30 dimensions. How are the Benign and Malignant cell readings distributed across the dimensions? In other words what is different between Benign and Malignant cells? Visualising the data on parallel coordinates gives a sense of its dimensions. (Red=Malignant Blue=Benign)

 As an example the perimeter dimension in the top chart is interesting. Filtering based on that looks like this.
While the perimeter dimension cannot help with a binary decision, visualisation shows that the compactness and concavity dimensions can help to decide from that point on. When you go through the above visualization and interact with it, notice that it looks more like decision making. So is this data set a good candidate for Random Forest Classifier or Nearest Neighbour Classifier or SVC?

2) Prediction results

Applying these classifiers to out of bag test data yields different levels of accuracy in prediction. Random Forest Classifier is more accurate than Nearest Neighbour Classifier. The result of applying these classifiers is shown below. RandomForestClassifier/GradientBoostingClassifier is the best choice as was clear from the visualization,

3) Confirm dimension insights from visualization
Fitted tree models have an attribute that shows the weight of each dimension. This allows us to confirm if the dimensions we thought were important from the visualization are really of any predictive value.

From the two plots notice that the dimensions radius, perimeter, concavity, concave_points have more weights than others in both classifiers. Notice that the standard error dimensions are less important.

Tuesday, 23 January 2018

Multi-DB design for high performance web applications

Applications with a small user base employ a design with a single centralised database. The rate of data growth and load do not create contention for such applications. So a single db design will work well. However, each database has a maximum connection limit. PostgreSQL is 'good for a few hundred connections, but for 1000s it would be better to look at a connection pooling solution'. Caching and db connection pooling can be utilised to postpone hitting the limit. This single db approach is easy to maintain, test and orchestrate with developer operations. 

When there are specific performance requirements for #users to support, #api requests per second, read/write patterns and data isolation the single database design would be broken by design. Hitting its upper limit on performance/load will impact users. A competition that has a better design to address this will win. A multi-database design addresses the shortcomings by separating data across multiple databases.

A) Data can be logically partitioned. For example, user authentication and profile information can reside on one database while other application data like say catalog, reviews and feedbacks can reside in a different database. That way authentication connections are routed to that specific server while the other servers can utilise connections fully to provision data.

B) Another approach is to partition based on read and write operations. All writes will go to one database. This database will be periodically synced to another read-only database. This helps when there are regular writes to the db but, the read density is too high. Multiple read-only slaves which mirror the primary can be used as needed.

C) A third approach is to use dedicated databases for different regions/locations. This is useful when usage varies across locations. Each database can be configured differently to handle its own access/load patterns. For example, utilise separate databases to store data from each state, New South Wales, Victoria, WA and so on.

D) To partition all data with redundancy, it would be better to use a database that has sharding built into its design. MongoDB is a good choice. This is described in a previous post.

Multiple databases mean more effort in development, operations, support and maintenance. So tangible performance benefits must justify their use.

Friday, 29 December 2017

Australia Tax Stats 2011-12 | Visualisation & Insights

This post visualises Australian Tax data for Individuals during the income year 2011-12. This web application allows the user to explore tax filing data to derive a number of insights. The data is available from the Australia Government website here

Watch the video for a look at this application. 

Specific Insights

1) Size of the workforce has not changed much: If the Wikipedia page on Australian economy is correct, then the workforce has not changed from 2011-12. The workforce count on the data confirms this.

2) Workforce made up of Equal numbers from both genders: Tax data shows that almost the same number of men and women figure in the filings.

3) State-wise gender gap in registered salary: Data shows that women, when compared to men, although roughly the same number in the workforce as men have their cumulative salary field at half the $ number in WA. This gap varies for the states. Since the number of women is almost the same, this wage gap triggers cause-analysis such as:

  1. Women are paid less in the same job compared to men.
  2. If (1) is not true then, women are not in the same pay band/scale or jobs as men are in. This would explain why the salary is less even though women make up the same head count.
4) Taxable Income/Salaries for under 18 age and above 70: The data shows that there is healthy income (but < 1 Billion) for these two age groups.

5) Comparison of states: On all accounts the ranking is NSW, Victoria, Queensland, WA, SA, TAS and other.

There are other data points like child support by filing individual, refundable, exemptions etc which can be analysed.

Thursday, 14 December 2017

Data visualisation: Plotting USGS earthquakes

This post makes use of USGS's live feed of earthquakes. The feed is plotted on a Leaflet map to visualize the data. Leaflet is a javascript library for interactive maps ( The advantage is that it consumes Geojson well.  

Feeds are categorised into all quakes and significant ones. The app's hourly feed is updated continuously for new data. When new data is available the map is updated. For each point of interest a leaflet layer group is utilised. This holds a circular marker layer and a place holder. A popup layer is also added which displays a text with the title from geojson. This title includes magnitude and location of the quake. All layer groups are added to the map's layer group. This allows individual layers associated with a point to be update or removed when data is updated.

Although all earth quakes are not significant, from the data is clear that earthquakes are more common in plate boundaries and fault lines (not surprising but nice to visualise). Switching to a weekly or monthly view makes vulnerable locations apparent.

Significant quakes are also seen in daily, weekly and monthly basis. Adding filters to query on Magnitude or changing the circle marker circumference based on magnitude would be good. However, significant earthquakes are far less in number and are shown separately in the app. Also to be noted is that, leaflet slowed down when a lot of layers were used for each point to support individual point updates. It also eases out when the map is just cleared and updates are added.


Python 3.5.3
Django 1.10
Bootstrap 3.3.7
Leaflet 1.2.0
Javascript, JQuery

Wednesday, 1 November 2017

Building a recommendation engine for Dota2

This post builds a recommendation engine that suggests heroes for a Dota2 player. This was one in a 3 set requirements to be completed. For more about Dota2 read on wikipedia and this. Two other requirements include player comparison and a leader board.

The application that was built looks like this.

Requirement: Dota2 has a set of 113 heroes and a huge community of players. The requirement was to suggest a hero for a given player. The recommendation engine should take into account play history of the specific player and heroes. Dota2 open API documentation is available at and the application must use this as a source of data for recommendations. The API for the solution must be available as REST and also on a web page. 

Approach: The solution is built as a collaborative item-to-item recommender. The data set for training and building the recommender is available from this api. An example for this api call to dota2 looks like this The result of this query has the hero play history (30 days back) for the player with id 87568060. For a list of players this can be utilised to build a data set. The advantage of using this api query is that it fits the recommender well as it has the play history for the heroes (by the players). And by using the data for all the players it contributes to the collaborative nature of the recommender. The resulting data set has the following columns and rows for each player to hero game plays.

player_id -> steam id for a player  
hero_id  -> id of the hero
plays -> number of times that this player played the hero.

This information is used to build a co-occurence matrix of all heroes to the player's heroes. This matrix is normalised using Jaccard index and a sorted weighted sum is applied to get the list of recommendations. Dota2 open api allows only 3 requests per second. So memcached is used to cache results of the external dota2 api call. The data set is divided into training and test respectively. The matrix is built on the training data. The matrix is built as follows

The disadvantage of this approach is that there needs to be a play history in place. Otherwise the matrix would be zero. This is the same for a new hero. One approach is to suggest a set of easy to play heroes to beginners. Again, this implementation builds the matrix for a subset of the players or for the whole community. The first case can be modified to build the matrix for only play history of beginners which would take into account hero plays of other beginners (collaborative).

Tools used:

Python 3.5.3
Django 1.10 & Django REST Framework
Pandas 0.20.3
Postgresql 9.5


For a good basic introduction to recommender systems see

Tuesday, 17 October 2017

Django models: I will cache you if I can

This post presents a generalised approach to making cache keys for Django model from fields and related one-to-one fields. In the last post  a Django model based approach to cache keys was utilised. However, that results in code duplication for the purpose of generating cache keys. At first look that appears inevitable since each model is cached in its own way. However as more requirements come in, a pattern can be identified and code duplication can be avoided. This new approach to making cache keys is enabled by a base class that inherits from models.Model and provides mechanisms to generate keys. The cache key generation takes into account what model fields the key has to be based on. 

For example, if there is a Player model with fields id, email, username. We may want a cache key that would look like '<the_id>'. Same for email and username. However, what if many fields of a model need to be in the cache key? This post presents a solution. Again, what if there is a one-to-one model which has a foreign key back to Player and that model needs to be cached ? For example, consider a model PlayerStats with fields id, user, stats. When this model needs to be queried it is likely to be based on the user. So, it makes sense to have a cache key for PlayerStats based on its linked player. Since this approach takes in any field, this becomes easy. The base class is

The class methods give back the cache key template and the instance method generates the cache key for the instance. The template is needed when some model's fields+values are available and we need to check the cache for those. Example is a web request asking for User with id=10 and we want to check that if that user is in cache.

The Player model will inherit from CacheableModel,

The PlayerStats looks like this

Now all that remains is a utility that takes in Model classes and calls the two class/instance methods.

- Now when Player needs to be queried on id, say from a Django view it becomes as easy as.

model_ins_from_cache_by_fields(Player, {'id': steamid})

- For Player stats for a specific player, this becomes

player = model_ins_from_cache_by_fields(Player, {'id': steamid})
player_totals = model_ins_from_cache_by_fields(PlayerTotals, {'player': player})

That is saving a lot of "if in cache else" code from repeating in views.

Some things to keep in mind:

1) Notice that the __str__ of the models is used for building keys. So make sure that __str__ of models do not have spaces and control characters.

2) Since it is based on __str__, if the __str__ represention is too long then the cache key becomes too long, possibly to the extent of increasing the time of hashing for the key.

Both these are addressable and is left as homework for reader. ;-)

Wednesday, 30 August 2017

Tuning the word cloud: Nuances with Mobile-support/REST-api/Cache

The word cloud web application in the previous post has been updated with the following features.

1) Mobile/Tablet support using Bootstrap css.
2) Configurable ignore lists to reduce noise in the cloud.
3) Caching.
4) REST.

1)  Mobile/Tablet support

Bootstrap makes the app usable on various device screens. Bootstrap intro page here. A page from the application now looks like this.

2) Ignore list via Configuration 

A web page can now be set with a list of words to ignore. There is no point so far in counting the word "like" on a social media page. Now the application can have words associated with lists and those lists associated with a web page. i.e configurable ignore lists. Example.

After applying an ignore list there is visible reduction in noise. (before Vs after).

3) Caching

This is most challenging in any application. When the data has not changed in the db, time spent on db/disk accesses for the same can be saved thus improving app latency. Memcached is used on 2 hosts and the application is configured to use the same. When a page is requested, data access is first checked against the cache, if it is a miss then only the db is queried. This data from the db is cached right away so that subsequent requests for the same data return fast. Common challenges in caching is described below. It is important to develop a feature without caching and then put the caching logic in. Screens show cache hits/miss on 2 consecutive requests for the same data. The second request returns from cache hits.


Data needs to be provisioned in a way that can be consumed by mobile applications, web apps or any program that wants to talk to your application. REST is good especially for native mobile applications that need the data but, handle display on their own. Django REST makes this manageable with a few quirks of its own. A rest response from the application using Chrome is shown and the same is accessible using curl command too.

The rest of the post is about common programming challenges in DRF and caching.

DRF challenge

DRF is good at so many levels. However it is tightly couple with the queryset. So if there is a list of instances rather than a queryset, things become incompatible. For example, customizing a foreign key related field to use cached data before hitting the database. While this may be a remote requirement the problem is well explained here. Curiously, Django1.10 has a new feature that lets queryset api with instances list. This api feature is described here. So if the instances that match are already in the cache they can be used. DRF serializer can be made to work with model instances like

The eager loading idea is from here but has been modified for the specific requirement with instances. After this what remains is overriding the get_queryset on the rest view. Works with Django 1.10. If the api is heavily used and database accesses can be saved, it is worth the effort and work too.

Caching Challenges


1) What to cache? 

To avoid access to the database, it is common to cache database entries. Images, static pages and json responses are also cached.


2) Cache code

Where to put it? How generalized should it be? Good cache key? In tutorials it is common to see cache access code and if that fails the object is retrieved from db and set to cache. This is leads to a lot of duplication of code. It is better to identify common sql access patterns to the db and cache those. The code that caches the db can be set as a behavior of the database entity or coded as a utility. Example for cache access logic as part of the db entity facade/ORM. This orm approach is described here.

Modify that for reuse: common accesses like the above on primary key can be factored out for re-use in a utility module.

Some access types are difficult to generalize and doing so can make the code base difficult to maintain. For example, fetching an entry in table A that corresponds to a foreign key. While it may be tempting (and feasible) to make a utility for this too, it is better to leave this alone until replication makes the use-cases clear and creates a demand for factoring this out into a utility.
What works in terms of generalizing the cache code depends on how many requirements come up. After a number of use cases, a general pattern should emerge in any application, like the primary key and foreign key access above.


3) Cache Key? 

A good cache key is one which describes the cached content. Also it should not be so long that the time to compute the hash beats the whole purpose.


4) Cache invalidation? 

If the data that is cached has changed then the cache needs to be updated. If the data is updated at many points in the application, it may become difficult to keep track of this. The example shows a signal handler that updates the cache on a db update/save. This approach works well.


5) Cache loading

Finally, when the application starts up, it is good to have the cache initiated with some data. What qualifies to be loaded to cache on startup? A fixed number of entries/tables with fixed row count or most frequented data points make good candidates. The application data, UI and usage need to be analyzed to determine the same. For example, data that corresponds to drop down lists with fixed number of entries. Say the names of states in a country shown in a drop down list.