# Most strongly related columns in the startup funding data The strongest relationships involve the descriptive text columns, especially SubVertical, Industry Vertical and Remarks. Many of the top scores are probably inflated by how many distinct values those columns have. The result is a normalized mutual information (NMI) ranking, which measures any kind of association, not only linear correlation. It comes from the step-1 result, and 1.0 means one column fully determines the other. **Top pairs by NMI** - **SubVertical and year (0.974):** SubVertical is about 82% unique (uniqueness score 0.82). Almost every value appears once, so it nearly "identifies" a row and therefore its year. This score is mostly an artifact. - **Industry Vertical and Remarks (0.888):** Remarks has few distinct values (uniqueness score 0.02), and the two columns overlap heavily. This may be a real link, but I haven't checked what the Remarks values are. - **Investment type and SubVertical (0.769), city and SubVertical (0.694), amount bucket and SubVertical (0.655):** These are again driven by SubVertical's very high cardinality. - **Industry Vertical and year (0.582), Remarks and SubVertical (0.526), Industry Vertical and SubVertical (0.525):** These are moderate, and the same cardinality caveat applies. **Likely more reliable associations** These involve columns with fewer distinct values, so the scores are less inflated. - **City and Industry Vertical (0.425):** Certain cities specialise in certain sectors. - **Industry Vertical and Investment type (0.419):** The funding round type differs by sector. - **Amount bucket and Investment type (0.267):** Deal size tracks round type, as you'd expect. - **Investment type and year (0.247), city and year (0.226):** There is some shift in round mix and city mix over time. **Weakest links** - **Amount and year (0.036):** Ticket size barely varies by year. - **Amount and city (0.070):** Ticket size barely varies by city. - **Amount and Remarks (0.112):** Ticket size is only weakly tied to Remarks. **Other caveats** - Startup Name and Investors Name are 100% unique, so they carry no usable association with anything. - Amount in USD is stored as text (VARCHAR), so the amount comparisons rely on a log10 bucketing of the values. A different bucketing could change those scores. If you want, I can recompute the associations after grouping rare SubVertical values, to see which relationships hold up without the cardinality effect.