The goal of this project is to develop a machine learning pipeline which is able to extract relevant information from provided invoice scans and predict a spending category for the expenses. A dataset of invoices is provided by the electoral college and the target categories are suggested by Dr. Power and his team. Unfortunately, the target categories are only available for the accumulated expenses between a party and a supplier, therefore, the amounts for the individual invoices have to be determined before the classification of invoices can be trained/evaluated.
The initial challenge arrises from the fact that there's no direct link between the invoices provided by the electoral college and the categories assigned to the expenses by Dr. Power and his team. The invoice list obtained from the electoral college, lists every invoice with an unique ID, the party, the supplier, a rough estimation of the expense category, the expense, and an ID if there is an attachement provided (e.g. scan of the physical invoice). The coding list provided by Dr. Power lists for every existing combination of party and suppliers a detailed breakdown which amounts were spent on each of the 58 listed categories. However, this can be an accumulation of multiple invoices between the party and a particular supplier and hence, this bridge must be gaped in order to obtain a supervised training set, where each invoice can be linked to one or multiple spending categories.
Before solving the task of matching individual invoice positions, we first aimed to match the total amounts. By summing up all expenses between each unique party - supplier combination in both the invoice and the coding list, only 50% of the amounts could be matched. Misspellings, white spaces, and different spellings of both, parties and suppliers, were identified as issue (e.g. GOOGLE UK and Google). This problem is tackled in connect_csv by removing whitespaces and keywords like "Ltd" from names. There is also a long list of converting names to unifying pendants which was developed during multiple iterations of identifying similar names or similar unlinked amounts (connect_amounts.ipynb) as identified in the snippet below.
After all adjustments, £49,540,640.20 (98.9%) could be matched, whereas £20,115.23 (0.0%) of the amounts listed in the coding list could not be linked to the invoices and £516,562.91 (1.0%) of the amounts in the invoices could not be linked to the coding list.
The goal now is to match the coded categorisations to the individual invoices. Since there can be multiple different categories in a coding of many invoices, bruteforcing a solution might become computationally unfeasable. Therefore heuristics such as greedy or genetic algorithms for this variation of the Multiple Knapsack problem should be considered.
Current implementations include:
- Random solver
- amounts matched: ~£44,000,000 (~88%)
- Genetic algorithm
- amounts matched: ~£49,500,000 (~99%)
Note that amounts matched does not equal amounts correctly matched as misassignments in one invoice can be belanced by misassignments in another invoice between the same supplier - party. Currently, the solvers can only assign one type of expense to each invoice although invoices can contain expenses from multiple categories. The next step would be identifying the matches with the biggest errors (and provided invoice copies) and manually reading them. The results should be entered in 'data/input/invoices_manual.csv' and the invoice matching updated accordingly.
The matched invoices can be found in invoices_matched.csv
The script to download the invoice files takes the invoice list obtained from the electoral college, identifies every entry with a provided supporting invoice, and uses these IDs to download the provided files. In total, 22,720 positions are listed by the electoral college, out of which 6,396 have an invoice attached.
To be continued.
Dr Sam Power, Lecturer in Corruption Analysis in the Politics Department, University of Sussex
On what services is money spent at elections? This is primary question behind this project. We know that just over £50 million was spent at the last general election, but strikingly little about how. Parties have to report their spending to the Electoral Commission under broad categories (e.g. ‘advertising’ and ‘market research and canvassing’), but this provides very little detail. They do, however, have to provide invoices for any spend over £200 so there is a vast resource available to find out more.
To address this gap, I (with a research team) manually coded every invoice (6,396) that was submitted to the Electoral Commission by political parties at the 2019 general election – which allowed us to paint a better picture of overall spend. These findings were written up into a report, a blog, and covered in the national media. The challenge here might seem obvious. It took ages and was incredibly arduous. There is much more information, available from many other elections, that could be analysed. I think, quite simply, there must be a better way. And perhaps that is via machine learning.
What we have to facilitate this better way, is a fantastic training set. I have provided two excel spreadsheets which contain the spending from the coding exercise. They have the name of the party conducting the spend, the amount of spend and the category that said spend falls under (you can see the amount of money written in the category). The Electoral Commission ‘political finance database’ hosts all of the invoices.
The suppliers that political parties use at elections took on new prominence after the Cambridge Analytica Scandal. And concerns about the ways in which modern technology, particularly social media platforms, are changing elections – and more widely democracy itself – are front and centre in the minds of many of us that study politics. But without a clear understanding on just what money is spent on, we cannot make a reasonable judgement about whether these trends are damaging and, if they are, how to combat them.